Large language models (LLMs) are typically aligned to be harmless to humans. Unfortunately, recent work has shown that such models are susceptible to automated jailbreak attacks that induce them to generate harmful content. More recent LLMs often incorporate an additional layer of defense, a Guard Model, which is a second LLM that is designed to check and moderate the output response of the primary LLM. Our key contribution is to show a novel attack strategy, PRP, that is successful against several open-source (e.g., Llama 2) and closed-source (e.g., GPT 3.5) implementations of Guard Models. PRP leverages a two step prefix-based attack that operates by (a) constructing a universal adversarial prefix for the Guard Model, and (b) propagating this prefix to the response. We find that this procedure is effective across multiple threat models, including ones in which the adversary has no access to the Guard Model at all. Our work suggests that further advances are required on defenses and Guard Models before they can be considered effective.
Large Language Models' success on text generation has also made them better at code generation and coding tasks. While a lot of work has demonstrated their remarkable performance on tasks such as code completion and editing, it is still unclear as to why. We help bridge this gap by exploring to what degree auto-regressive models understand the logical constructs of the underlying programs. We propose Counterfactual Analysis for Programming Concept Predicates (CACP) as a counterfactual testing framework to evaluate whether Large Code Models understand programming concepts. With only black-box access to the model, we use CACP to evaluate ten popular Large Code Models for four different programming concepts. Our findings suggest that current models lack understanding of concepts such as data flow and control flow.
Adversarial examples threaten the integrity of machine learning systems with alarming success rates even under constrained black-box conditions. Stateful defenses have emerged as an effective countermeasure, detecting potential attacks by maintaining a buffer of recent queries and detecting new queries that are too similar. However, these defenses fundamentally pose a trade-off between attack detection and false positive rates, and this trade-off is typically optimized by hand-picking feature extractors and similarity thresholds that empirically work well. There is little current understanding as to the formal limits of this trade-off and the exact properties of the feature extractors/underlying problem domain that influence it. This work aims to address this gap by offering a theoretical characterization of the trade-off between detection and false positive rates for stateful defenses. We provide upper bounds for detection rates of a general class of feature extractors and analyze the impact of this trade-off on the convergence of black-box attacks. We then support our theoretical findings with empirical evaluations across multiple datasets and stateful defenses.
Defending machine-learning (ML) models against white-box adversarial attacks has proven to be extremely difficult. Instead, recent work has proposed stateful defenses in an attempt to defend against a more restricted black-box attacker. These defenses operate by tracking a history of incoming model queries, and rejecting those that are suspiciously similar. The current state-of-the-art stateful defense Blacklight was proposed at USENIX Security '22 and claims to prevent nearly 100% of attacks on both the CIFAR10 and ImageNet datasets. In this paper, we observe that an attacker can significantly reduce the accuracy of a Blacklight-protected classifier (e.g., from 82.2% to 6.4% on CIFAR10) by simply adjusting the parameters of an existing black-box attack. Motivated by this surprising observation, since existing attacks were evaluated by the Blacklight authors, we provide a systematization of stateful defenses to understand why existing stateful defense models fail. Finally, we propose a stronger evaluation strategy for stateful defenses comprised of adaptive score and hard-label based black-box attacks. We use these attacks to successfully reduce even reconfigured versions of Blacklight to as low as 0% robust accuracy.
Voice assistants are deployed widely and provide useful functionality. However, recent work has shown that commercial systems like Amazon Alexa and Google Home are vulnerable to voice-based confusion attacks that exploit design issues. We propose a systems-oriented defense against this class of attacks and demonstrate its functionality for Amazon Alexa. We ensure that only the skills a user intends execute in response to voice commands. Our key insight is that we can interpret a user's intentions by analyzing their activity on counterpart systems of the web and smartphones. For example, the Lyft ride-sharing Alexa skill has an Android app and a website. Our work shows how information from counterpart apps can help reduce dis-ambiguities in the skill invocation process. We build SkilIFence, a browser extension that existing voice assistant users can install to ensure that only legitimate skills run in response to their commands. Using real user data from MTurk (N = 116) and experimental trials involving synthetic and organic speech, we show that SkillFence provides a balance between usability and security by securing 90.83% of skills that a user will need with a False acceptance rate of 19.83%.
Content scanning systems employ perceptual hashing algorithms to scan user content for illegal material, such as child pornography or terrorist recruitment flyers. Perceptual hashing algorithms help determine whether two images are visually similar while preserving the privacy of the input images. Several efforts from industry and academia propose to conduct content scanning on client devices such as smartphones due to the impending roll out of end-to-end encryption that will make server-side content scanning difficult. However, these proposals have met with strong criticism because of the potential for the technology to be misused and re-purposed. Our work informs this conversation by experimentally characterizing the potential for one type of misuse -- attackers manipulating the content scanning system to perform physical surveillance on target locations. Our contributions are threefold: (1) we offer a definition of physical surveillance in the context of client-side image scanning systems; (2) we experimentally characterize this risk and create a surveillance algorithm that achieves physical surveillance rates of >40% by poisoning 5% of the perceptual hash database; (3) we experimentally study the trade-off between the robustness of client-side image scanning systems and surveillance, showing that more robust detection of illegal material leads to increased potential for physical surveillance.
Detecting deepfakes is an important problem, but recent work has shown that DNN-based deepfake detectors are brittle against adversarial deepfakes, in which an adversary adds imperceptible perturbations to a deepfake to evade detection. In this work, we show that a modification to the detection strategy in which we replace a single classifier with a carefully chosen ensemble, in which input transformations for each model in the ensemble induces pairwise orthogonal gradients, can significantly improve robustness beyond the de facto solution of adversarial training. We present theoretical results to show that such orthogonal gradients can help thwart a first-order adversary by reducing the dimensionality of the input subspace in which adversarial deepfakes lie. We validate the results empirically by instantiating and evaluating a randomized version of such "orthogonal" ensembles for adversarial deepfake detection and find that these randomized ensembles exhibit significantly higher robustness as deepfake detectors compared to state-of-the-art deepfake detectors against adversarial deepfakes, even those created using strong PGD-500 attacks.
Physical adversarial examples for camera-based computer vision have so far been achieved through visible artifacts -- a sticker on a Stop sign, colorful borders around eyeglasses or a 3D printed object with a colorful texture. An implicit assumption here is that the perturbations must be visible so that a camera can sense them. By contrast, we contribute a procedure to generate, for the first time, physical adversarial examples that are invisible to human eyes. Rather than modifying the victim object with visible artifacts, we modify light that illuminates the object. We demonstrate how an attacker can craft a modulated light signal that adversarially illuminates a scene and causes targeted misclassifications on a state-of-the-art ImageNet deep learning model. Concretely, we exploit the radiometric rolling shutter effect in commodity cameras to create precise striping patterns that appear on images. To human eyes, it appears like the object is illuminated, but the camera creates an image with stripes that will cause ML models to output the attacker-desired classification. We conduct a range of simulation and physical experiments with LEDs, demonstrating targeted attack rates up to 84%.