Image watermarking embeds imperceptible signals into AI-generated images for deepfake detection and provenance verification. Although recent semantic-level watermarking methods demonstrate strong resistance against conventional pixel-level removal attacks, their robustness against more advanced removal strategies remains underexplored, raising concerns about their reliability in practical scenarios. Existing removal attacks primarily operate in the pixel domain without altering image semantics, which limits their effectiveness against semantic-level watermarks. In this paper, we propose Next Frame Prediction Attack (NFPA), the first semantic-level removal attack. Unlike pixel-level attacks, NFPA formulates watermark removal as a video generation task: it treats the watermarked image as the initial frame and aims to subtly manipulate the image semantics to generate the next-frame image, i.e., the unwatermarked image. We conduct a comprehensive evaluation on eight state-of-the-art image watermarking schemes, demonstrating that NFPA consistently outperforms thirteen removal attack baselines in terms of the trade-off between watermark removal and image quality. Our results reveal the vulnerabilities of current image watermarking methods and highlight the urgent need for more robust watermarks.

All current backdoor attacks on deep learning (DL) models fall under the category of a vertical class backdoor (VCB).In VCB attacks, any sample from a class activates the implanted backdoor when the secret trigger is present, regardless of whether it is a sub-type source-class-agnostic backdoor or a source-class-specific backdoor. For example, a trigger of sunglasses could mislead a facial recognition model when either an arbitrary (source-class-agnostic) or a specific (source-class-specific) person wears sunglasses. Existing defense strategiesoverwhelmingly focus on countering VCB attacks, especially those that are source-class-agnostic. This narrow focus neglects the potential threat of other simpler yet general backdoor types, leading to false security implications. It is, therefore, crucial to discover and elucidate unknown backdoor types, particularly those that can be easily implemented, as a mandatory step before developing countermeasures.

Deep Neural Network (DNN) models are often deployed in resource-sharing clouds as Machine Learning as a Service (MLaaS) to provide inference services. To steal model architectures that are of valuable intellectual properties, a class of attacks has been proposed via different side-channel leakage, posing a serious security challenge to MLaaS.Also targeting MLaaS, we propose a new end-to-end attack, DeepTheft, to accurately recover complex DNN model architectures on general processors via the RAPL (Running Average Power Limit)-based power side channel. While unprivileged access to the RAPL has been disabled in bare-metal OSes, we observe that the RAPL is still legitimately accessible in a platform as a service, e.g., the latest docker environment of version 20.10.18 used in this work. However, an attacker can acquire only a low sampling rate (1 KHz) of the time-series energy traces from the RAPL interface, rendering existing techniques ineffective in stealing large and deep DNN models. To this end, we design a novel and generic learning-based framework consisting of a set of meta-models, based on which DeepTheft is demonstrated to have high accuracy in recovering a large number (thousands) of models architectures from different model families including the deepest ResNet152. Particularly, DeepTheft has achieved a Levenshtein Distance Accuracy of 99.75% in recovering network structures, and a weighted average F1 score of 99.60% in recovering diverse layer-wise hyperparameters. Besides, our proposed learning framework is general to other time-series side-channel signals. To validate its generalization, another existing side channel is exploited, i.e., CPU frequency. Different from RAPL, CPU frequency is accessible to unprivileged users in bare-metal OSes. By using our generic learning framework trained against CPU frequency traces, DeepTheft has shown similarly high attack performance in stealing model architectures.

Deep neural networks (DNNs) are susceptible to backdoor attacks, where malicious functionality is embedded to allow attackers to trigger incorrect classifications. Old-school backdoor attacks use strong trigger features that can easily be learned by victim models. Despite robustness against input variation, the robustness however increases the likelihood of unintentional trigger activations. This leaves traces to existing defenses, which find approximate replacements for the original triggers that can activate the backdoor without being identical to the original trigger via, e.g., reverse engineering and sample overlay.In this paper, we propose and investigate a new characteristic of backdoor attacks, namely, backdoor exclusivity, which measures the ability of backdoor triggers to remain effective in the presence of input variation. Building upon the concept of backdoor exclusivity, we propose Backdoor Exclusivity LifTing (BELT), a novel technique which suppresses the association between the backdoor and fuzzy triggers to enhance backdoor exclusivity for defense evasion. Extensive evaluation on three popular backdoor benchmarks validate, our approach substantially enhances the stealthiness of four old-school backdoor attacks, which, after backdoor exclusivity lifting, is able to evade seven state-of-the-art backdoor countermeasures, at almost no cost of the attack success rate and normal utility. For example, one of the earliest backdoor attacks BadNet, enhanced by BELT, evades most of the state-of-the-art defenses including ABS and MOTH which would otherwise recognize the backdoored model.

Due to their low latency and high privacy preservation, there is currently a burgeoning demand for deploying deep learning (DL) models on ubiquitous edge Internet of Things (IoT) devices. However, DL models are often large in size and require large-scale computation, which prevents them from being placed directly onto IoT devices, where resources are constrained, and 32-bit floating-point (float-32) operations are unavailable. Commercial framework (i.e., a set of toolkits) empowered model quantization is a pragmatic solution that enables DL deployment on mobile devices and embedded systems by effortlessly post-quantizing a large high-precision model (e.g., float-32) into a small low-precision model (e.g., int-8) while retaining the model inference accuracy. However, their usability might be threatened by security vulnerabilities. This work reveals that standard quantization toolkits can be abused to activate a backdoor. We demonstrate that a full-precision backdoored model which does not have any backdoor effect in the presence of a trigger—as the backdoor is dormant—can be activated by (i) TensorFlow-Lite (TFLite) quantization, the onlyproduct-readyquantization framework to date, and (ii) thebeta releasedPyTorch Mobile framework. In our experiments, we employ three popular model architectures (VGG16, ResNet18, and ResNet50), and train each across three popular datasets: MNIST, CIFAR10 and GTSRB. We ascertain that all trained float-32 backdoored models exhibit no backdoor effecteven in the presence of trigger inputs. Particularly, four influential backdoor defenses are evaluated, and they fail to identify a backdoor in the float-32 models. When each of the float-32 models is converted into an int-8 format model through the standard TFLite or PyTorch Mobile framework's post-training quantization, the backdoor is activated in the quantized model, which shows a stable attack success rate close to 100% upon inputs with the trigger, while it usually behaves upon non-trigger inputs. This work highlights that a stealthy security threat occurs when an end-user utilizes the on-device post-training model quantization frameworks, informing security researchers of a cross-platform overhaul of DL models post-quantization even if these models pass security-aware front-end backdoor inspections. Significantly, we have identified Gaussian noise injection into the malicious full-precision model as an easy-to-use preventative defense against the PQ backdoor. The attack source code is released athttps://github.com/quantization-backdoor.

Deep learning models with backdoors act maliciously when triggered but seem normal otherwise. This risk, often increased by model outsourcing, challenges their secure use. Although countermeasures exist, their defense against adaptive attacks is under-examined, possibly leading to security misjudgments. This study is the first intricate examination illustrating the difficulty of detecting backdoors in outsourced models, especially when attackers adjust their strategies, even if their capabilities are significantly limited. It is relatively straightforward for attackers to circumvent detection by trivially violating its threat model (e.g., using advanced backdoor types or trigger designs not covered by the detection). However, this research highlights that various leading detection defenses can simultaneously be evaded using simple adaptive strategies, even under their defined threat models and with limited adversary capabilities (e.g., using easily detectable triggers while maintaining a high attack success rate). To be more specific, this study introduces a novel methodology that employs trigger specificity enhancement and training regulation in a symbiotic manner. This approach allows us to evade multiple backdoor detection defenses simultaneously, including Neural Cleanse (Oakland 19’), ABS (CCS 19’), and MNTD (Oakland 21’). These were the detection tools selected for the Evasive Trojans Track of the 2022 NeurIPS Trojan Detection Challenge. Even when applied in conjunction with these defenses under stringent conditions, such as a high attack success rate (> 97%) and the restricted use of the simplest trigger (small white square), our straightforward method garnered the second prize in NeurIPS Trojan Detection Challenge. Notably, for the first time, our adaptive attack successfully evaded other recent state-of-the-art defenses, including FeatureRE (NeurIPS 22’) and Beatrix (NDSS 23’). This study suggests that existing model outsourcing backdoor defenses remain vulnerable to adaptive attacks, and thus, the use of third-party models should be avoided whenever possible.

Since Deep Learning (DL) backdoor attacks have been revealed as one of the most insidious adversarial attacks, a number of countermeasures have been developed with certain assumptions defined in their respective threat models. However, their robustness is currently inadvertently ignored, which can introduce severe consequences, e.g., a countermeasure can be misused and result in a false implication of backdoor detection. For the first time, we critically examine the robustness of existing backdoor countermeasures. As an initial study, we first identify five potential non-robust failure factors including binary classification, poison rate, model complexity, single-model justification, and hyperparameter sensitivity. As exhaustively examining defenses is infeasible, we instead focus on influential backdoor detection-based countermeasures consisting of model-inspection ones including Neural Cleanse (S&P’19), ABS (CCS’19), and MNTD (S&P’21), and data-inspection ones including SCAn (USENIX SECURITY’21) to examine their failure cases under one or more of these factors. Although these investigated countermeasures claim that they work well under their respective threat models, they have inherent unexplored non-robust cases, which are not even rooted from delicate adaptive attacks. We demonstrate how to trivially bypass them aligned with their respective threat models by simply varying the aforementioned factors. Particularly, for each defense, formal proofs or empirical studies are used to reveal its non-robust cases where it is not as robust as it claims or expects. This work highlights the necessity of thoroughly evaluating the robustness of backdoor countermeasures to avoid their misleading security implications in unknown non-robust cases.

Currently, a high demand for on-device deep neural network (DNN) model deployment is limited by the large model size, computing-intensive floating-point operations (FLOPS), and intellectual property (IP) infringements (i.e., easy access to model duplication for the avoidance of license payments). One appealing solution to addressing the first two concerns is model quantization, which reduces the model size and uses integer operations commonly supported by microcontrollers (MCUs usually do not support FLOPS). To this end, a 1-bit quantized DNN model or deep binary neural network (BNN) significantly improves the memory efficiency, where each parameter in a BNN model has only 1 bit. However, BNN cannot directly provide IP protection (in particular, the functionality of the model is locked unless there is a license payment). In this article, we propose a reconfigurable BNN (RBNN) to further amplify the memory efficiency for resource-constrained Internet of Things (IoT) devices while naturally protecting the model IP. Generally, RBNN can be reconfigured on demand to achieve any one of$M$($M>1$) distinct tasks with the same parameter set, thus only a single task determines the memory requirements. In other words, the memory utilization is improved by a factor of$M$. Our extensive experiments corroborate that up to seven commonly used tasks ($M=7$, six of these tasks are image related and the last one is audio) can co-exist (the value of$M$can be larger). These tasks with a varying number of classes have no or negligible accuracy drop-off (i.e., within 1%) on three binarized popular DNN architectures, including VGG, ResNet, and ReActNet. The tasks span across different domains, e.g., computer vision and audio domains validated herein, with the prerequisite that the model architecture can serve those cross-domain tasks. To fulfill the IP protection of an RBNN model, the reconfiguration can be controlled by both a user key and a device-unique root key generated by the intrinsic hardware fingerprint (e.g., SRAM memory power-up pattern). By doing so, an RBNN model can only be used per paid user per authorized device, thus benefiting both the user and the model provider. The source code is released athttps://github.com/LearningMaker/RBNN.