Viet Q. Vo

Despite the significant efforts devoted to aligning large language models (LLMs) with human values and ensuring safe deployment, recent work has revealed that LLMs remain vulnerable to adversarial jailbreak attacks that can bypass safety guardrails and elicit harmful responses. Many defense methods are proposed to detect jailbreaks but they are limited in their effectiveness to counter wide-range optimization based jailbreak mechanisms that can yield highly fluency-optimized or harmful semantic obfuscated prompts. To tackle this challenge, we propose a unified detection framework SAFEGuard—which incorporates a hybrid fluency measurement based on cross-layer distribution distance and perplexity, and the analysis of harmful semantics through gradient matching. Our method is grounded in a paramount observation: high fluency prompts maintain their malicious intention close to harmful prompts while harmful semantic obfuscated prompts often inject gibberish token sequences. Our evaluation demonstrates that SAFEGuard consistently outperforms state-of-the-art baselines and achieves significant improvement in accuracy across different optimization-based jailbreaks. This underscores the effectiveness of SAFEGuard against evolving jailbreak attacks.

Certified defenses promise provable robustness guarantees. We study the malicious exploitation of probabilistic certification frameworks to better understand the limits of guarantee provisions. Now, the objective is to not only mislead a classifier, but also manipulate the certification process to generate a robustness guarantee for an adversarial input certificate spoofing. A recent study in ICLR demonstrated that crafting large perturbations can shift inputs far into regions capable of generating a certificate for an incorrect class. Our study investigates if perturbations needed to cause a misclassification and yet coax a certified model into issuing a deceptive, large robustness radius for a target class can still be made small and imperceptible. We explore the idea of region-focused adversarial examples to craft imperceptible perturbations, spoof certificates and achieve certification radii larger than the source class ghost certificates. Extensive evaluations with the ImageNet demonstrate the ability to effectively bypass state-of-the-art certified defenses such as Densepure. Our work underscores the need to better understand the limits of robustness certification methods.

We study the unique, less-well understood problem of generating sparse adversarial samples simply by observing the score-based replies to model queries. Sparse attacks aim to discover a minimum number—the bounded—perturbations to model inputs to craft adversarial examples and misguide model decisions. But, in contrast to query-based dense attack counterparts against black-box models, constructing sparse adversarial perturbations, even when models serve confidence score information to queries in a score-based setting, is non-trivial. Because, such an attack leads to: i) an NP-hard problem; and ii) a non-differentiable search space. We develop the BRUSLEATTACK—a new, faster (more query-efficient) algorithm formulation for the problem. We conduct extensive attack evaluations including an attack demonstration against a Machine Learning as a Service (MLaaS) offering exemplified by Google Cloud Vision and robustness testing of adversarial training regimes and a recent defense against black-box attacks. The proposed attack scales to achieve state-of-the-art attack success rates and query efficiency on standard computer vision tasks such as ImageNet across different model architectures. Our artifacts and DIY attack samples are available on GitHub. Importantly, our work facilitates faster evaluation of model vulnerabilities and raises our vigilance on the safety, security and reliability of deployed systems.

Despite our best efforts, deep learning models remain highly vulnerable to even tiny adversarial perturbations applied to the inputs. The ability to extract information for solely the output of a machine learning model to craft adversarial perturbations to black-box models is a practical threat against real-world systems, such as autonomous cars or machine learning models exposed as a service (MLaaS). Of particular interest are sparse attacks. The realization of sparse attacks in black-box models demonstrates that machine learning models are more vulnerable than we believe. Because, these attacks aim to minimize the number of perturbed pixels—measured byl0norm—required to mislead a model by solely observing the decision (the predicted label) returned to a model query; the so-called decision-based attack setting. But, such an attack leads to an NP-hard optimization problem. We develop an evolution-based algorithm—SparseEvo—for the problem and evaluate against both convolutional deep neural networks and vision transformers. Notably, vision transformers are yet to be investigated under a decision-based at-tack setting. SparseEvo requires significantly fewer model queries than the state-of-the-art sparse attack Pointwise for both untargeted and targeted attacks. The attack algorithm, although conceptually simple, is also competitive with only a limited query budget against the state-of-the-art gradient-based Whitebox attacks in standard computer vision tasks such as ImageNet. Importantly, the query efficient SparseEvo, along with decision-based attacks, in general, raise new questions regarding the safety of deployed systems and poses new directions to study and understand the robustness of machine learning models.

Machine learning models are critically susceptible to evasion attacks from adversarial examples. Generally, adversarial examples, modified inputs deceptively similar to the original input, are constructed under whitebox settings by adversaries with full access to the model. However, recent attacks have shown a remarkable reduction in query numbers to craft adversarial examples using blackbox attacks. Particularly, alarming is the ability to exploit the classification decision from the access interface of a trained model provided by a growing number of Machine Learning as a Service providers including Google, Microsoft, IBM and used by a plethora of applications incorporating these models. The ability of an adversary to exploit only the predicted label from a model to craft adversarial examples is distinguished as a decision-based attack. In our study, we first deep dive into recent state-of-the-art decision-based attacks in ICLR and SP to highlight the costly nature of discovering low distortion adversarial employing gradient estimation methods. We develop a robust query efficient attack capable of avoiding entrapment in a local minimum and misdirection from noisy gradients seen in gradient estimation methods. The attack method we propose, RamBoAttack, exploits the notion of Randomized Block Coordinate Descent to explore the hidden classifier manifold, targeting perturbations to manipulate only localized input features to address the issues of gradient estimation methods. Importantly, the RamBoAttack is more robust to the different sample inputs available to an adversary and the targeted class. Overall, for a given target class, RamBoAttack is demonstrated to be more robust at achieving a lower distortion within a given query budget. We curate our extensive results using the large-scale high-resolution ImageNet dataset and open-source our attack, test samples and artifacts on GitHub.