We evaluate the effectiveness of filtering child images from training datasets of text-to-image models to prevent model misuse to create child sexual abuse material (CSAM). First, we capture the complexity of preventing CSAM generation using a game-based security definition. Second, we show that current detection methods cannot remove all children from a dataset. Third, using an ethical proxy for CSAM (a child wearing glasses), we show that even when only a small percentage of child images are left in the training dataset after filtering, there exist prompting strategies that generate a child wearing glasses using only a few more queries than when the model is trained on the unfiltered data. Fine-tuning the filtered model on child images further reduces the additional query overhead. We also show that re-introducing a concept is possible via fine-tuning even if filtering is perfect. Our results show that current child filtering methods offer limited protection to closed-weight models and no protection to open-weight models, while reducing the generality of the model by hindering the generation of child-related concepts or changing their representation. We conclude by outlining challenges in conducting evaluations that establish robust evidence on the impact of concept filtering defenses for CSAM.

Query-based systems (QBSs) are one of the key approaches for sharing data. QBSs allow analysts to request aggregate information from a private protected dataset. Attacks are a crucial part of ensuring QBSs are truly privacy-preserving. The development and testing of attacks is however very labor-intensive and unable to cope with the increasing complexity of systems. Automated approaches have been shown to be promising but are currently extremely computationally intensive, limiting their applicability in practice. We here propose QueryCheetah, a fast and effective method for automated discovery of privacy attacks against QBSs. We instantiate QueryCheetah on attribute inference attacks and show it to discover stronger attacks than previous methods while being 18 times faster than the state-of-the-art automated approach. We then show how QueryCheetah allows system developers to thoroughly evaluate the privacy risk, including for various attacker strengths and target individuals. We finally show how QueryCheetah can be used out-of-the-box to find attacks in larger syntaxes and workarounds around ad-hoc defenses.

Smart meters, devices measuring the electricity and gas consumption of a household, are currently being deployed at a fast rate throughout the world. The data they collect are extremely useful, including in the fight against climate change. However, these data and the information that can be inferred from them are highly sensitive. Re-pseudonymization, i.e., the frequent replacement of random identifiers over time, is widely used to share smart meter data while mitigating the risk of re-identification. We here show how, in spite of re-pseudonymization, households' consumption records can be pieced together with high accuracy in large-scale datasets. More specifically, we propose the first deep learning-based profiling attack against re-pseudonymized smart meter data. Our attack combines neural network embeddings, which are used to extract features from weekly consumption records and are tailored to the smart meter identification task, with a nearest neighbor classifier. We evaluate six neural networks architectures as the embedding model. Our results suggest that the Transformer and CNN-LSTM architectures vastly outperform previous methods as well as other architectures, successfully identifying the correct household 73.4% of the time among 5139 households based on electricity and gas consumption records (54.5% for electricity only). We further show that the features extracted by the embedding model maintain their effectiveness when transferred to a set of users disjoint from the one used to train the model. Finally, we extensively evaluate the robustness of our results. In particular, we show that the accuracy of the attack only slowly decreases with the size of the dataset, how less frequent re-pseudonymization will further increase accuracy, and how an attacker can evaluate the likelihood of a match to be correct. Taken together, our results strongly suggest that even frequent re-pseudonymization strategies can be reversed, strongly limiting their ability to prevent re-identification in practice.

Machine learning models have been shown to leak sensitive information about their training datasets. Models are increasingly deployed on devices, raising concerns that white-box access to the model parameters increases the attack surface compared to black-box access which only provides query access. Directly extending the shadow modelling technique from the black-box to the white-box setting has been shown, in general, not to perform better than black-box only attacks. A potential reason is misalignment, a known characteristic of deep neural networks. In the shadow modelling context, misalignment means that, while the shadow models learn similar features in each layer, the features are located in different positions. We here present the first systematic analysis of the causes of misalignment in shadow models and show the use of a different weight initialisation to be the main cause. We then extend several re-alignment techniques, previously developed in the model fusion literature, to the shadow modelling context, where the goal is to re-align the layers of a shadow model to those of the target model.We show re-alignment techniques to significantly reduce the measured misalignment between the target and shadow models. Finally, we perform a comprehensive evaluation of white-box membership inference attacks (MIA). Our analysis reveals that internal layer activation-based MIAs suffer strongly from shadow model misalignment, while gradient-based MIAs are only sometimes significantly affected. We show that re-aligning the shadow models strongly improves the former's performance and can also improve the latter's performance, although less frequently. On the CIFAR10 dataset with a false positive rate of 1%, white-box MIA using re-aligned shadow models improves the true positive rate by 4.5%.Taken together, our results highlight that on-device deployment increases the attack surface and that the newly available information can be used to build more powerful attacks.

Location data is frequently collected from populations and shared in aggregate form to guide policy and decision making. However, the prevalence of aggregated data also raises the privacy concern of membership inference attacks (MIAs). MIAs infer whether an individual's data contributed to the aggregate release. Although effective MIAs have been developed for aggregate location data, these require access to an extensive auxiliary dataset of individual traces over the same locations, which are collected from a similar population. This assumption is often impractical given common privacy practices surrounding location data. To measure the risk of an MIA performed by a realistic adversary, we develop the first Zero Auxiliary Knowledge (ZK) MIA on aggregate location data, which eliminates the need for an auxiliary dataset of real individual traces. Instead, we develop a novel synthetic approach, such that suitable synthetic traces are generated from the released aggregate. We also develop methods to correct for bias and noise, to show that our synthetic-based attack is still applicable when privacy mechanisms are applied prior to release. Using two large-scale location datasets, we demonstrate that our ZK MIA matches the state-of-the-art Knock-Knock (KK) MIA across a wide range of settings, including popular implementations of differential privacy (DP) and suppression of small counts. Furthermore, we show that ZK MIA remains highly effective even when the adversary only knows a small fraction (10%) of their target's location history. This demonstrates that effective MIAs can be performed by realistic adversaries, highlighting the need for strong DP protection.

End-to-end encryption (E2EE) provides strong technical protections to individuals from interferences. Governments and law enforcement agencies around the world have however raised concerns that E2EE also allows illegal content to be shared undetected. Client-side scanning (CSS), using perceptual hashing (PH) to detect known illegal content before it is shared, is seen as a promising solution to prevent the diffusion of illegal content while preserving encryption. While these proposals raise strong privacy concerns, proponents of the solutions have argued that the risk is limited as the technology has a limited scope: detecting known illegal content. In this paper, we show that modern perceptual hashing algorithms are actually fairly flexible pieces of technology and that this flexibility could be used by an adversary to add a secondary hidden feature to a client-side scanning system. More specifically, we show that an adversary providing the PH algorithm can "hide" a secondary purpose of face recognition of a target individual alongside its primary purpose of image copy detection. We first propose a procedure to train a dual-purpose deep perceptual hashing model by jointly optimizing for both the image copy detection and the targeted facial recognition task. Second, we extensively evaluate our dual-purpose model and show it to be able to reliably identify a target individual 67% of the time while not impacting its performance at detecting illegal content. We also show that our model is neither a general face detection nor a facial recognition model, allowing its secondary purpose to be hidden. Finally, we show that the secondary purpose can be enabled by adding a single illegal looking image to the database. Taken together, our results raise concerns that a deep perceptual hashing-based CSS system could turn billions of user devices into tools to locate targeted individuals.

Although query-based systems (QBS) have become one of the main solutions to share data anonymously, building QBSes that robustly protect the privacy of individuals contributing to the dataset is a hard problem. Theoretical solutions relying on differential privacy guarantees are difficult to implement correctly with reasonable accuracy, while ad-hoc solutions might contain unknown vulnerabilities. Evaluating the privacy provided by QBSes must thus be done by evaluating the accuracy of a wide range of privacy attacks. However, existing attacks against QBSes require time and expertise to develop, need to be manually tailored to the specific systems attacked, and are limited in scope. In this paper, we develop QuerySnout, the first method to automatically discover vulnerabilities in query-based systems. QuerySnout takes as input a target record and the QBS as a black box, analyzes its behavior on one or more datasets, and outputs a multiset of queries together with a rule to combine answers to them in order to reveal the sensitive attribute of the target record. QuerySnout uses evolutionary search techniques based on a novel mutation operator to find a multiset of queries susceptible to lead to an attack, and a machine learning classifier to infer the sensitive attribute from answers to the queries selected. We showcase the versatility of QuerySnout by applying it to two attack scenarios (assuming access to either the private dataset or to a different dataset from the same distribution), three real-world datasets, and a variety of protection mechanisms. We show the attacks found by QuerySnout to consistently equate or outperform, sometimes by a large margin, the best attacks from the literature. We finally show how QuerySnout can be extended to QBSes that require a budget, and apply QuerySnout to a simple QBS based on the Laplace mechanism. Taken together, our results show how powerful and accurate attacks against QBSes can already be found by an automated system, allowing for highly complex QBSes to be automatically tested "at the pressing of a button". We believe this line of research to be crucial to improve the robustness of systems providing privacy-preserving access to personal data in theory and in practice.

The Winograd Schema Challenge (WSC) dataset WSC273 and its inference counterpart WNLI are popular benchmarks for natural language understanding and commonsense reasoning.In this paper, we show that the performance of three language models on WSC273 consistently and robustly improves when finetuned on a similar pronoun disambiguation problem dataset (denoted WSCR).We additionally generate a large unsupervised WSClike dataset.By fine-tuning the BERT language model both on the introduced and on the WSCR dataset, we achieve overall accuracies of 72.5% and 74.7% on WSC273 and WNLI, improving the previous state-of-theart solutions by 8.8% and 9.6%, respectively.Furthermore, our fine-tuned models are also consistently more accurate on the "complex" subsets of WSC273, introduced by Trichelair et al. (2018).

Vid Kocijan, Oana-Maria Camburu, Ana-Maria Cretu, Yordan Yordanov, Phil Blunsom, Thomas Lukasiewicz. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.