State-of-the-art Diffusion Models (DMs) produce highly realistic images. While prior work has successfully mitigated Not Safe For Work (NSFW) content in the visual domain, we identify a novel threat: the generation of NSFW text embedded within images. This includes offensive language, such as insults, racial slurs, and sexually explicit terms, posing significant risks to users. We show that all state-of-the-art DMs (e.g., SD3, SDXL, Flux, DeepFloyd IF) are vulnerable to this issue. Through extensive experiments, we demonstrate that existing mitigation techniques, effective for visual content, fail to prevent harmful text generation while substantially degrading benign text generation. As an initial step toward addressing this threat, we introduce a novel fine-tuning strategy that targets only the text-generation layers in DMs. Therefore, we construct a safety fine-tuning dataset by pairing each NSFW prompt with two images: one with the NSFW term, and another where that term is replaced with a carefully crafted benign alternative while leaving the image unchanged otherwise. By training on this dataset, the model learns to avoid generating harmful text while preserving benign content and overall image quality. Finally, to advance research in the area, we release ToxicBench, an open-source benchmark for evaluating NSFW text generation in images. It includes our curated fine-tuning dataset, a set of harmful prompts, new evaluation metrics, and a pipeline that assesses both NSFW-ness and text and image quality. Our benchmark aims to guide future efforts in mitigating NSFW text generation in text-to-image models, thereby contributing to their safe deployment.

Current graph neural network (GNN) model-stealing methods rely heavily on queries to the victim model, assuming no hard query limits. However, in reality, the number of allowed queries can be severely limited. In this paper, we demonstrate how an adversary can extract a GNN with very limited interactions with the model. Our approach first enables the adversary to obtain the model backbone without making direct queries to the victim model and then to strategically utilize a fixed query limit to extract the most informative data. The experiments on eight real-world datasets demonstrate the effectiveness of the attack, even under a very restricted query limit and under defense against model extraction in place. Our findings underscore the need for robust defenses against GNN model extraction threats.

Graph Neural Networks (GNNs) achieve high performance in various real-world applications, such as drug discovery, traffic states prediction, and recommendation systems. The fact that building powerful GNNs requires a large amount of training data, powerful computing resources, and human expertise turns the models into lucrative targets for model stealing attacks. Prior work has revealed that the threat vector of stealing attacks against GNNs is large and diverse, as an attacker can leverage various heterogeneous signals ranging from node labels to high-dimensional node embeddings to create a local copy of the target GNN at a fraction of the original training costs. This diversity in the threat vector renders the design of effective and general defenses challenging and existing defenses usually focus on one particular stealing setup. Additionally, they solely provide means to identify stolen model copies rather than preventing the attack. To close this gap, we propose the first and general Active Defense Against GNN Extraction (ADAGE). ADAGE builds on the observation that stealing a model's full functionality requires highly diverse queries to leak its behavior across the input space. Our defense monitors this query diversity and progressively perturbs outputs as the accumulated leakage grows. In contrast to prior work, ADAGE can prevent stealing across all common attack setups. Our extensive experimental evaluation using six benchmark datasets, four GNN models, and three types of adaptive attackers shows that ADAGE penalizes attackers to the degree of rendering stealing impossible, whilst preserving predictive performance on downstream tasks. ADAGE, thereby, contributes towards securely sharing valuable GNNs in the future.

Modern image generative models are capable of producing photorealistic images that become increasingly indistinguishable from real data. As those images are published online, they are often scraped for subsequent training runs of new generative models. This practice of training on generated data has been shown to degrade model performance and amplify existing biases. A possible mitigation lies in embedding watermarks into generated content to identify synthetic content and enable data provenance. However, the persistence of watermarks through iterative training-generation cycles remains poorly understood. In this work, we investigate how image watermarks degrade across multiple generations of model training. We adopt a dataset inference framework that aggregates weak per-sample signals via statistical hypothesis testing, enabling reliable detection even when watermark traces are subtle. We observe disparate behavior across watermarking methods with one of them remaining highly detectable in subsequent generations, even when the watermarked training data constitutes only 1% of the training data, while other watermarks degrade severely throughout generations and have a finite lifetime. Our findings highlight the need for more robust watermarking methods that are designed to withstand retraining for reliable data provenance.

Machine learning (ML) models have been shown to leak private information from their training datasets. Differential Privacy (DP), typically implemented through the differential private stochastic gradient descent algorithm (DP-SGD), has become the standard solution to bound leakage from the models. Despite recent improvements, DP-SGD-based approaches for private learning still usually struggle in the high privacy (ε≤1) and low data regimes, and when the private training datasets are imbalanced. To overcome these limitations, we propose Differentially Private Prototype Learning (DPPL) as a new paradigm for private transfer learning. DPPL leverages publicly pre-trained encoders to extract features from private data and generates DP prototypes that represent each private class in the embedding space and can be publicly released for inference. Since our DP prototypes can be obtained from only a few private training data points and without iterative noise addition, they offer high-utility predictions and strong privacy guarantees even under the notion of pure DP. We additionally show that privacy-utility trade-offs can be further improved when leveraging the public data beyond pre-training of the encoder: in particular, we can privately sample our DP prototypes from the publicly available data points used to train the encoder. Our experimental evaluation with four state-of-the-art encoders, four vision datasets, and under different data and imbalancedness regimes demonstrate DPPL's high performance under strong privacy guarantees in challenging private learning setups.

Differentially private stochastic gradient descent (DP-SGD) trains machine learning (ML) models with formal privacy guarantees for the training set by adding random noise to gradient updates. In collaborative learning (CL), where multiple parties jointly train a model, noise addition occurs either (i) before or (ii) during secure gradient aggregation. The first option is deployed in distributed DP methods, which require greater amounts of total noise to achieve security, resulting in degraded model utility. The second approach preserves model utility but requires a secure multiparty computation (MPC) protocol. Existing methods for MPC noise generation require tens to hundreds of seconds of runtime per noise sample because of the number of parties involved. This makes them impractical for collaborative learning, which often requires thousands or more samples of noise in each training step.

Diffusion Models (DMs) benefit from large and diverse datasets for their training. Since this data is often scraped from the Internet without permission from the data owners, this raises concerns about copyright and intellectual property protections. While (illicit) use of data is easily detected for training samples perfectly re-created by a DM at inference time, it is much harder for data owners to verify if their data was used for training when the outputs from the suspect DM are not close replicas. Conceptually, membership inference attacks (MIAs), which detect if a given data point was used during training, present themselves as a suitable tool to address this challenge. However, we demonstrate that existing MIAs are not strong enough to reliably determine the membership of individual images in large, state-of-the-art DMs. To overcome this limitation, we propose Copyrighted Data Identification (CDI), a framework for data owners to identify whether their dataset was used to train a given DM. CDI relies on dataset inference techniques, i.e., instead of using the membership signal from a single data point, CDI leverages the fact that most data owners, such as providers of stock photography, visual media companies, or even individual artists, own datasets with multiple publicly exposed data points which might all be included in the training of a given DM. By selectively aggregating signals from existing MIAs and using new handcrafted methods to extract features from these datasets, feeding them to a scoring model, and applying rigorous statistical testing, CDI allows data owners with as little as 70 data points to identify with a confidence of more than 99% whether their data was used to train a given DM. Thereby, CDI represents a valuable tool for data owners to claim illegitimate use of their copyrighted data. We make our code available at https://github.com/sprintml/copyrighted_data_identification.

Federated learning (FL) with differential privacy (DP) provides a framework for collaborative machine learning, enabling clients to train a shared model while adhering to strict privacy constraints. The framework allows each client to have an individual privacy guarantee, e.g., by adding different amounts of noise to each client's model updates. One underlying assumption is that all clients spend their privacy budgets uniformly over time (learning rounds). However, it has been shown in the literature that learning in early rounds typically focuses on more coarse-grained features that can be learned at lower signal-to-noise ratios while later rounds learn fine-grained features that benefit from higher signal-to-noise ratios. Building on this intuition, we propose a time-adaptive DP-FL framework that expends the privacy budget non-uniformly across both time and clients. Our framework enables each client to save privacy budget in early rounds so as to be able to spend more in later rounds when additional accuracy is beneficial in learning more fine-grained features. We theoretically prove utility improvements in the case that clients with stricter privacy budgets spend budgets unevenly across rounds, compared to clients with more relaxed budgets, who have sufficient budgets to distribute their spend more evenly. Our practical experiments on standard benchmark datasets support our theoretical results and show that, in practice, our algorithms improve the privacy-utility trade-offs compared to baseline schemes.

Novel diffusion models can synthesize photo-realistic images with integrated high-quality text. Surprisingly, we demonstrate through attention activation patching that only less than $1$\% of diffusion models' parameters, all contained in attention layers, influence the generation of textual content within the images. Building on this observation, we improve textual generation efficiency and performance by targeting cross and joint attention layers of diffusion models. We introduce several applications that benefit from localizing the layers responsible for textual content generation. We first show that a LoRA-based fine-tuning solely of the localized layers enhances, even more, the general text-generation capabilities of large diffusion models while preserving the quality and diversity of the diffusion models' generations. Then, we demonstrate how we can use the localized layers to edit textual content in generated images. Finally, we extend this idea to the practical use case of preventing the generation of toxic text in a cost-free manner. In contrast to prior work, our localization approach is broadly applicable across various diffusion model architectures, including U-Net (e.g., SDXL and DeepFloyd IF) and transformer-based (e.g., Stable Diffusion 3), utilizing diverse text encoders (e.g., from CLIP to the large language models like T5). Project page available at https://t2i-text-loc.github.io/.

Multi-modal models, such as CLIP, have demonstrated strong performance in aligning visual and textual representations, excelling in tasks like image retrieval and zero-shot classification. Despite this success, the mechanisms by which these models utilize training data, particularly the role of memorization, remain unclear. In uni-modal models, both supervised and self-supervised, memorization has been shown to be essential for generalization. However, it is not well understood how these findings would apply to CLIP, which incorporates elements from both supervised learning via captions that provide a supervisory signal similar to labels, and from self-supervised learning via the contrastive objective. To bridge this gap in understanding, we propose a formal definition of memorization in CLIP (CLIPMem) and use it to quantify memorization in CLIP models. Our results indicate that CLIP’s memorization behavior falls between the supervised and self-supervised paradigms, with "mis-captioned" samples exhibiting highest levels of memorization. Additionally, we find that the text encoder contributes more to memorization than the image encoder, suggesting that mitigation strategies should focus on the text domain. Building on these insights, we propose multiple strategies to reduce memorization while at the same time improving utility---something that had not been shown before for traditional learning paradigms where reducing memorization typically results in utility decrease.

Image AutoRegressive generation has emerged as a new powerful paradigm with image autoregressive models (IARs) matching state-of-the-art diffusion models (DMs) in image quality (FID: 1.48 vs. 1.58) while allowing for a higher generation speed. However, the privacy risks associated with IARs remain unexplored, raising concerns regarding their responsible deployment. To address this gap, we conduct a comprehensive privacy analysis of IARs, comparing their privacy risks to the ones of DMs as reference points. Concretely, we develop a novel membership inference attack (MIA) that achieves a remarkably high success rate in detecting training images (with a True Positive Rate at False Positive Rate = 1% of 86.38% vs. 6.38% for DMs with comparable attacks). We leverage our novel MIA to provide dataset inference (DI) for IARs, and show that it requires as few as 6 samples to detect dataset membership (compared to 200 for DI in DMs), confirming a higher information leakage in IARs. Finally, we are able to extract hundreds of training data points from an IAR (e.g., 698 from VAR-d30). Our results suggest a fundamental privacy-utility trade-off: while IARs excel in image generation quality and speed, they are empirically significantly more vulnerable to privacy attacks compared to DMs that achieve similar performance. We release the code at https://github.com/sprintml/privacy_attacks_against_iars for reproducibility.

Prompting has become a dominant paradigm for adapting large language models (LLMs). While discrete (textual) prompts are widely used for their interpretability, soft (parameter) prompts have recently gained traction in APIs. This is because they can encode information from more training samples while minimizing the user’s token usage, leaving more space in the context window for task-specific input. However, soft prompts are tightly coupled to the LLM they are tuned on, limiting their generalization to other LLMs. This constraint is particularly problematic for efficiency and privacy: (1) tuning prompts on each LLM incurs high computational costs, especially as LLMs continue to grow in size. Additionally, (2) when the LLM is hosted externally, soft prompt tuning often requires sharing private data with the LLM provider. For instance, this is the case with the NVIDIA NeMo API. To address these issues, we propose POST (Privacy Of Soft prompt Transfer), a framework that enables private tuning of soft prompts on a small model and subsequently transfers these prompts to a larger LLM. POST uses knowledge distillation to derive a small model directly from the large LLM to improve prompt transferability, tunes the soft prompt locally, optionally with differential privacy guarantees, and transfers it back to the larger LLM using a small public dataset. Our experiments show that POST reduces computational costs, preserves privacy, and effectively transfers high-utility soft prompts.

The remarkable capabilities of Large Language Models (LLMs) can be mainly attributed to their massive training datasets, which are often scraped from the internet without respecting data owners’ intellectual property rights. Dataset Inference (DI) offers a potential remedy by identifying whether a suspect dataset was used in training, thereby enabling data owners to verify unauthorized use. However, existing DI methods require a private set—known to be absent from training—that closely matches the compromised dataset’s distribution. Such in-distribution, held-out data is rarely available in practice, severely limiting the applicability of DI. In this work, we address this challenge by synthetically generating the required held-out set. Our approach tackles two key obstacles: (1) creating high-quality, diverse synthetic data that accurately reflects the original distribution, which we achieve via a data generator trained on a carefully designed suffix-based completion task, and (2) bridging likelihood gaps between real and synthetic data, which is realized through post-hoc calibration. Extensive experiments on diverse text datasets show that using our generated data as a held-out set enables DI to detect the original training sets with high confidence, while maintaining a low false positive rate. This result empowers copyright owners to make legitimate claims on data usage and demonstrates our method’s reliability for real-world litigations. Our code is available at https://github.com/sprintml/PostHocDatasetInference.

State-of-the-art membership inference attacks (MIAs) typically require training many reference models, making it difficult to scale these attacks to large pre-trained language models (LLMs). As a result, prior research has either relied on weaker attacks that avoid training references (e.g., fine-tuning attacks), or on stronger attacks applied to small models and datasets. However, weaker attacks have been shown to be brittle and insights from strong attacks in simplified settings do not translate to today's LLMs. These challenges prompt an important question: are the limitations observed in prior work due to attack design choices, or are MIAs fundamentally ineffective on LLMs? We address this question by scaling LiRA--one of the strongest MIAs--to GPT-2 architectures ranging from 10M to 1B parameters, training references on over 20B tokens from the C4 dataset. Our results advance the understanding of MIAs on LLMs in four key ways. While (1) strong MIAs can succeed on pre-trained LLMs, (2) their effectiveness, remains limited (e.g., AUC<0.7) in practical settings. (3) Even when strong MIAs achieve better-than-random AUC, aggregate metrics can conceal substantial per-sample MIA decision instability: due to training randomness, many decisions are so unstable that they are statistically indistinguishable from a coin flip. Finally, (4) the relationship between MIA success and related LLM privacy metrics is not as straightforward as prior work has suggested.

Deep neural networks (DNNs) have been shown to memorize their training data, but similar analyses for graph neural networks (GNNs) remain under-explored. We introduce NCMemo (Node Classification Memorization), the first framework to quantify label memorization in semi-supervised node classification. We establish an inverse relationship between memorization and graph homophily, i.e the tendency of connected nodes to share labels or features. Lower homophily significantly increases memorization, indicating that GNNs rely on label memorization when learning less homophilic graphs. We then analyze GNN training dynamics and find that increased memorization in low-homophily graphs is tightly coupled to GNNs' implicit bias toward using graph structure. When structure is less informative, models instead memorize node labels to minimize training loss. Finally, we show that nodes with higher label inconsistency in their feature-space neighborhood are more prone to memorization. Based on these insights, we investigate graph rewiring as a mitigation strategy. Our results show that rewiring reduces memorization without harming model performance, while also lowering the privacy risk for previously memorized data points. Thus, our work advances understanding of GNN learning and supports more privacy-preserving GNN deployment.

State-of-the-art text-to-image models generate photorealistic images at an unprecedented speed. This work focuses on models that operate in a bitwise autoregressive manner over a discrete set of tokens that is practically infinite in size. However, their impressive generative power comes with a growing risk: as their outputs increasingly populate the Internet, they are likely to be scraped and reused as training data—potentially by the very same models. This phenomenon has been shown to lead to model collapse, where repeated training on generated content, especially from the models’ own previous versions, causes a gradual degradation in performance. A promising mitigation strategy is watermarking, which embeds human-imperceptible yet detectable signals into generated images—enabling the identification of generated content. In this work, we introduce BitMark, a robust bitwise watermarking framework. Our method embeds a watermark directly at the bit level of the token stream during the image generation process. Our bitwise watermark subtly influences the bits to preserve visual fidelity and generation speed while remaining robust against a spectrum of removal techniques. Furthermore, it exhibits high radioactivity, i.e., when watermarked generated images are used to train another image generative model, this second model’s outputs will also carry the watermark. The radioactive traces remain detectable even when only fine-tuning diffusion or image autoregressive models on images watermarked with our BitMark. Overall, our approach provides a principled step toward preventing model collapse in image generative models by enabling reliable detection of generated outputs. The code is available at https://github.com/sprintml/BitMark.

Graph Neural Networks (GNNs) are recognized as potent tools for processing real-world data organized in graph structures. Especially inductive GNNs, which allow for the processing of graph-structured data without relying on predefined graph structures, are becoming increasingly important in a wide range of applications. As such these networks become attractive targets for model-stealing attacks where an adversary seeks to replicate the functionality of the targeted network. Significant efforts have been devoted to developing model-stealing attacks that extract models trained on images and texts. However, little attention has been given to stealing GNNs trained on graph data. This paper identifies a new method of performing unsupervised model-stealing attacks against inductive GNNs, utilizing graph contrastive learning and spectral graph augmentations to efficiently extract information from the targeted model. The new type of attack is thoroughly evaluated on six datasets and the results show that our approach outperforms the current state-of-the-art by Shen et al. (2021). In particular, our attack surpasses the baseline across all benchmarks, attaining superior fidelity and downstream accuracy of the stolen model while necessitating fewer queries directed toward the target model.

Self-supervised learning (SSL) has recently received significant attention due to its ability to train high-performance encoders purely on unlabeled data---often scraped from the internet. This data can still be sensitive and empirical evidence suggests that SSL encoders memorize private information of their training data and can disclose them at inference time. Since existing theoretical definitions of memorization from supervised learning rely on labels, they do not transfer to SSL. To address this gap, we propose a framework for defining memorization within the context of SSL. Our definition compares the difference in alignment of representations for data points and their augmented views returned by both encoders that were trained on these data points and encoders that were not. Through comprehensive empirical analysis on diverse encoder architectures and datasets we highlight that even though SSL relies on large datasets and strong augmentations---both known in supervised learning as regularization techniques that reduce overfitting---still significant fractions of training data points experience high memorization. Through our empirical results, we show that this memorization is essential for encoders to achieve higher generalization performance on different downstream tasks.

While open Large Language Models (LLMs) have made significant progress, they still fall short of matching the performance of their closed, proprietary counterparts, making the latter attractive even for the use on highly *private* data. Recently, various new methods have been proposed to adapt closed LLMs to private data without leaking private information to third parties and/or the LLM provider. In this work, we analyze the privacy protection and performance of the four most recent methods for private adaptation of closed LLMs. By examining their threat models and thoroughly comparing their performance under different privacy levels according to differential privacy (DP), various LLM architectures, and multiple datasets for classification and generation tasks, we find that: (1) all the methods leak query data, i.e., the (potentially sensitive) user data that is queried at inference time, to the LLM provider, (2) three out of four methods also leak large fractions of private training data to the LLM provider while the method that protects private data requires a local open LLM, (3) all the methods exhibit lower performance compared to three private gradient-based adaptation methods for *local open LLMs*, and (4) the private adaptation methods for closed LLMs incur higher monetary training and query costs than running the alternative methods on local open LLMs. This yields the conclusion that, to achieve truly *privacy-preserving LLM adaptations* that yield high performance and more privacy at lower costs, taking into account current methods and models, one should use open LLMs.

Diffusion models (DMs) produce very detailed and high-quality images. Their power results from extensive training on large amounts of data - usually scraped from the internet without proper attribution or consent from content creators. Unfortunately, this practice raises privacy and intellectual property concerns, as DMs can memorize and later reproduce their potentially sensitive or copyrighted training images at inference time. Prior efforts prevent this issue by either changing the input to the diffusion process, thereby preventing the DM from generating memorized samples during inference, or removing the memorized data from training altogether. While those are viable solutions when the DM is developed and deployed in a secure and constantly monitored environment, they hold the risk of adversaries circumventing the safeguards and are not effective when the DM itself is publicly released. To solve the problem, we introduce NeMo, the first method to localize memorization of individual data samples down to the level of neurons in DMs' cross-attention layers. Through our experiments, we make the intriguing finding that in many cases, single neurons are responsible for memorizing particular training samples. By deactivating these memorization neurons, we can avoid the replication of training data at inference time, increase the diversity in the generated outputs, and mitigate the leakage of private and copyrighted data. In this way, our NeMo contributes to a more responsible deployment of DMs.

The proliferation of large language models (LLMs) in the real world has come with a rise in copyright cases against companies for training their models on unlicensed data from the internet. Recent works have presented methods to identify if individual text sequences were members of the model's training data, known as membership inference attacks (MIAs). We demonstrate that the apparent success of these MIAs is confounded by selecting non-members (text sequences not used for training) belonging to a different distribution from the members (e.g., temporally shifted recent Wikipedia articles compared with ones used to train the model). This distribution shift makes membership inference appear successful. However, most MIA methods perform no better than random guessing when discriminating between members and non-members from the same distribution (e.g., in this case, the same period of time). Even when MIAs work, we find that different MIAs succeed at inferring membership of samples from different distributions. Instead, we propose a new dataset inference method to accurately identify the datasets used to train large language models. This paradigm sits realistically in the modern-day copyright landscape, where authors claim that an LLM is trained over multiple documents (such as a book) written by them, rather than one particular paragraph. While dataset inference shares many of the challenges of membership inference, we solve it by selectively combining the MIAs that provide positive signal for a given distribution, and aggregating them to perform a statistical test on a given dataset. Our approach successfully distinguishes the train and test sets of different subsets of the Pile with statistically significant p-values $< 0.1$, without any false positives.

Recent work on studying memorization in self-supervised learning (SSL) suggests that even though SSL encoders are trained on millions of images, they still memorize individual data points. While effort has been put into characterizing the memorized data and linking encoder memorization to downstream utility, little is known about where the memorization happens inside SSL encoders. To close this gap, we propose two metrics for localizing memorization in SSL encoders on a per-layer (LayerMem) and per-unit basis (UnitMem). Our localization methods are independent of the downstream task, do not require any label information, and can be performed in a forward pass. By localizing memorization in various encoder architectures (convolutional and transformer-based) trained on diverse datasets with contrastive and non-contrastive SSL frameworks, we find that (1) while SSL memorization increases with layer depth, highly memorizing units are distributed across the entire encoder, (2) a significant fraction of units in SSL encoders experiences surprisingly high memorization of individual data points, which is in contrast to models trained under supervision, (3) atypical (or outlier) data points cause much higher layer and unit memorization than standard data points, and (4) in vision transformers, most memorization happens in the fully-connected layers. Finally, we show that localizing memorization in SSL has the potential to improve fine-tuning and to inform pruning strategies.

In federated learning (FL), data does not leave personal devices when they are jointly training a machine learning model. Instead, these devices share gradients, parameters, or other model updates, with a central party (e.g., a company) coordinating the training. Because data never "leaves" personal devices, FL is often presented as privacy-preserving. Yet, recently it was shown that this protection is but a thin facade, as even a passive, honest-but-curious attacker observing gradients can reconstruct data of individual users contributing to the protocol.In this work, we show a novel data reconstruction attack which allows an active and dishonest central party to efficiently extract user data from the received gradients. While prior work on data reconstruction in FL relies on solving computationally expensive optimization problems or on making easily detectable modifications to the shared model’s architecture or parameters, in our attack the central party makes inconspicuous changes to the shared model’s weights before sending them out to the users. We call the modified weights of our attack trap weights.Our active attacker is able to recover user data perfectly, i.e., with zero error, even when this data stems from the same class. Recovery comes with near-zero costs: the attack requires no complex optimization objectives. Instead, our attacker exploits inherent data leakage from model gradients and simply amplifies this effect by maliciously altering the weights of the shared model through the trap weights. These specificities enable our attack to scale to fully-connected and convolutional deep neural networks trained with large mini-batches of data. For example, for the high-dimensional vision dataset ImageNet, we perfectly reconstruct more than 50% of the training data points from mini-batches as large as 100 data points. In textual tasks, such as IMDB sentiment analysis, more than 65% of data points from mini-batches containing 100 data points can be perfectly reconstructed.

Federated learning (FL) is a framework for users to jointly train a machine learning model. FL is promoted as a privacy-enhancing technology (PET) that provides data minimization: data never "leaves" personal devices and users share only model updates with a server (e.g., a company) coordinating the distributed training. While prior work showed that in vanilla FL a malicious server can extract users’ private data from the model updates, in this work we take it further and demonstrate that a malicious server can reconstruct user data even in hardened versions of the protocol. More precisely, we propose an attack against FL protected with distributed differential privacy (DDP) and secure aggregation (SA). Our attack method is based on the introduction of sybil devices that deviate from the protocol to expose individual users’ data for reconstruction by the server. The underlying root cause for the vulnerability to our attack is a power imbalance: the server orchestrates the whole protocol and users are given little guarantees about the selection of other users participating in the protocol. Moving forward, we discuss requirements for privacy guarantees in FL. We conclude that users should only participate in the protocol when they trust the server or they apply local primitives such as local DP, shifting power away from the server. Yet, the latter approaches come at significant overhead in terms of performance degradation of the trained model, making them less likely to be deployed in practice.

When training a machine learning model with differential privacy, one sets a privacy budget. This uniform budget represents an overall maximal privacy violation that any user is willing to face by contributing their data to the training set. We argue that this approach is limited because different users may have different privacy expectations. Thus, setting a uniform privacy budget across all points may be overly conservative for some users or, conversely, not sufficiently protective for others. In this paper, we capture these preferences through individualized privacy budgets. To demonstrate their practicality, we introduce a variant of Differentially Private Stochastic Gradient Descent (DP-SGD) which supports such individualized budgets. DP-SGD is the canonical approach to training models with differential privacy. We modify its data sampling and gradient noising mechanisms to arrive at our approach, which we call Individualized DP-SGD (IDP-SGD). Because IDP-SGD provides privacy guarantees tailored to the preferences of individual users and their data points, we empirically find it to improve privacy-utility trade-offs.

Large language models (LLMs) are excellent in-context learners. However, the sensitivity of data contained in prompts raises privacy concerns. Our work first shows that these concerns are valid: we instantiate a simple but highly effective membership inference attack against the data used to prompt LLMs. To address this vulnerability, one could forego prompting and resort to fine-tuning LLMs with known algorithms for private gradient descent. However, this comes at the expense of the practicality and efficiency offered by prompting. Therefore, we propose to privately learn to prompt. We first show that soft prompts can be obtained privately through gradient descent on downstream data. However, this is not the case for discrete prompts. Thus, we orchestrate a noisy vote among an ensemble of LLMs presented with different prompts, i.e., a flock of stochastic parrots. The vote privately transfers the flock's knowledge into a single public prompt. We show that LLMs prompted with our private algorithms closely match the non-private baselines. For example, using GPT3 as the base model, we achieve a downstream accuracy of 92.7% on the sst2 dataset with $(\varepsilon=0.147, \delta=10^{-6})$-differential privacy vs. 95.2% for the non-private baseline. Through our experiments, we also show that our prompt-based approach is easily deployed with existing commercial~APIs.

Machine Learning as a Service (MLaaS) APIs provide ready-to-use and high-utility encoders that generate vector representations for given inputs. Since these encoders are very costly to train, they become lucrative targets for model stealing attacks during which an adversary leverages query access to the API to replicate the encoder locally at a fraction of the original training costs. We propose *Bucks for Buckets (B4B)*, the first *active defense* that prevents stealing while the attack is happening without degrading representation quality for legitimate API users. Our defense relies on the observation that the representations returned to adversaries who try to steal the encoder's functionality cover a significantly larger fraction of the embedding space than representations of legitimate users who utilize the encoder to solve a particular downstream task. B4B leverages this to adaptively adjust the utility of the returned representations according to a user's coverage of the embedding space. To prevent adaptive adversaries from eluding our defense by simply creating multiple user accounts (sybils), B4B also individually transforms each user's representations. This prevents the adversary from directly aggregating representations over multiple accounts to create their stolen encoder copy. Our active defense opens a new path towards securely sharing and democratizing encoders over public APIs.

Collaborative machine learning (ML) is widely used to enable institutions to learn better models from distributed data. While collaborative approaches to learning intuitively protect user data, they remain vulnerable to either the server, the clients, or both, deviating from the protocol. Indeed, because the protocol is asymmetric, a malicious server can abuse its power to reconstruct client data points. Conversely, malicious clients can corrupt learning with malicious updates. Thus, both clients and servers require a guarantee when the other cannot be trusted to fully cooperate. In this work, we propose a peer-to-peer (P2P) learning scheme that is secure against malicious servers and robust to malicious clients. Our core contribution is a generic framework that transforms any (compatible) algorithm for robust aggregation of model updates to the setting where servers and clients can act maliciously. Finally, we demonstrate the computational efficiency of our approach even with 1-million parameter models trained by 100s of peers on standard datasets.

Applying machine learning (ML) to sensitive domains requires privacy protection of the underlying training data through formal privacy frameworks, such as differential privacy (DP). Yet, usually, the privacy of the training data comes at the cost of the resulting ML models' utility. One reason for this is that DP uses one uniform privacy budget epsilon for all training data points, which has to align with the strictest privacy requirement encountered among all data holders. In practice, different data holders have different privacy requirements and data points of data holders with lower requirements can contribute more information to the training process of the ML models. To account for this need, we propose two novel methods based on the Private Aggregation of Teacher Ensembles (PATE) framework to support the training of ML models with individualized privacy guarantees. We formally describe the methods, provide a theoretical analysis of their privacy bounds, and experimentally evaluate their effect on the final model's utility using the MNIST, SVHN, and Adult income datasets. Our empirical results show that the individualized privacy methods yield ML models of higher accuracy than the non-individualized baseline. Thereby, we improve the privacy-utility trade-off in scenarios in which different data holders consent to contribute their sensitive data at different individual privacy levels.

Private multi-winner voting is the task of revealing k-hot binary vectors satisfying a bounded differential privacy (DP) guarantee. This task has been understudied in machine learning literature despite its prevalence in many domains such as healthcare. We propose three new DP multi-winner mechanisms: Binary, Tau, and Powerset voting. Binary voting operates independently per label through composition. Tau voting bounds votes optimally in their L2 norm for tight data-independent guarantees. Powerset voting operates over the entire binary vector by viewing the possible outcomes as a power set. Our theoretical and empirical analysis shows that Binary voting can be a competitive mechanism on many tasks unless there are strong correlations between labels, in which case Powerset voting outperforms it. We use our mechanisms to enable privacy-preserving multi-label learning in the central setting by extending the canonical single-label technique: PATE. We find that our techniques outperform current state-of-the-art approaches on large, real-world healthcare data and standard multi-label benchmarks. We further enable multi-label confidential and private collaborative (CaPC) learning and show that model performance can be significantly improved in the multi-site setting.

Self-Supervised Learning (SSL) is an increasingly popular ML paradigm that trains models to transform complex inputs into representations without relying on explicit labels. These representations encode similarity structures that enable efficient learning of multiple downstream tasks. Recently, ML-as-a-Service providers have commenced offering trained SSL models over inference APIs, which transform user inputs into useful representations for a fee. However, the high cost involved to train these models and their exposure over APIs both make black-box extraction a realistic security threat. We thus explore model stealing attacks against SSL. Unlike traditional model extraction on classifiers that output labels, the victim models here output representations; these representations are of significantly higher dimensionality compared to the low-dimensional prediction scores output by classifiers. We construct several novel attacks and find that approaches that train directly on a victim’s stolen representations are query efficient and enable high accuracy for downstream models. We then show that existing defenses against model extraction are inadequate and not easily retrofitted to the specificities of SSL.

Self-supervised models are increasingly prevalent in machine learning (ML) since they reduce the need for expensively labeled data. Because of their versatility in downstream applications, they are increasingly used as a service exposed via public APIs. At the same time, these encoder models are particularly vulnerable to model stealing attacks due to the high dimensionality of vector representations they output. Yet, encoders remain undefended: existing mitigation strategies for stealing attacks focus on supervised learning. We introduce a new dataset inference defense, which uses the private training set of the victim encoder model to attribute its ownership in the event of stealing. The intuition is that the log-likelihood of an encoder's output representations is higher on the victim's training data than on test data if it is stolen from the victim, but not if it is independently trained. We compute this log-likelihood using density estimation models. As part of our evaluation, we also propose measuring the fidelity of stolen encoders and quantifying the effectiveness of the theft detection without involving downstream tasks; instead, we leverage mutual information and distance measurements. Our extensive empirical results in the vision domain demonstrate that dataset inference is a promising direction for defending self-supervised models against model stealing.

Although pretrained Transformers such as BERT achieve high accuracy on indistribution examples, do they generalize to new distributions?We systematically measure out-of-distribution (OOD) generalization for seven NLP datasets by constructing a new robustness benchmark with realistic distribution shifts.We measure the generalization of previous models including bag-of-words models, ConvNets, and LSTMs, and we show that pretrained Transformers' performance declines are substantially smaller.Pretrained transformers are also more effective at detecting anomalous or OOD examples, while many previous models are frequently worse than chance.We examine which factors affect robustness, finding that larger models are not necessarily more robust, distillation can be harmful, and more diverse pretraining data can enhance robustness.Finally, we show where future work can improve OOD robustness.

The convolutional layers are core building blocks of neural network architectures. In general, a convolutional filter applies to the entire frequency spectrum of the input data. We explore artificially constraining the frequency spectra of these filters and data, called band-limiting, during training. The frequency domain constraints apply to both the feed-forward and back-propagation steps. Experimentally, we observe that Convolutional Neural Networks (CNNs) are resilient to this compression scheme and results suggest that CNNs learn to leverage lower-frequency components. In particular, we found: (1) band-limited training can effectively control the resource usage (GPU and memory); (2) models trained with band-limited layers retain high prediction accuracy; and (3) requires no modification to existing training algorithms or neural network architectures to use unlike other compression schemes.

Commercial DBMSs, such as Microsoft SQL Server, cater to diverse workloads including transaction processing, decision support, and operational analytics. They also support variety in physical design structures such as B+ tree and columnstore. The benefits of B+ tree for OLTP workloads and columnstore for decision support workloads are well-understood. However, the importance of hybrid physical designs, consisting of both columnstore and B+ tree indexes on the same database, is not well-studied --- a focus of this paper. We first quantify the trade-offs using carefully-crafted micro-benchmarks. This micro-benchmarking indicates that hybrid physical designs can result in orders of magnitude better performance depending on the workload. For complex real-world applications, choosing an appropriate combination of columnstore and B+ tree indexes for a database workload is challenging. We extend the Database Engine Tuning Advisor for Microsoft SQL Server to recommend a suitable combination of B+ tree and columnstore indexes for a given workload. Through extensive experiments using industry-standard benchmarks and several real-world customer workloads, we quantify how a physical design tool capable of recommending hybrid physical designs can result in orders of magnitude better execution costs compared to approaches that rely either on columnstore-only or B+ tree-only designs.