Jacopo Cortellazzi

Yannick Assogba, Jacopo Cortellazzi, Javier Abad et al.

Jailbreak attacks remain a persistent threat to large language model safety. We propose Context-Conditioned Delta Steering (CC-Delta), an SAE-based defense that identifies jailbreak-relevant sparse features by comparing token-level representations of the same harmful request with and without jailbreak context. Using paired harmful/jailbreak prompts, CC-Delta selects features via statistical testing and applies inference-time mean-shift steering in SAE latent space. Across four aligned instruction-tuned models and twelve jailbreak attacks, CC-Delta achieves comparable or better safety-utility tradeoffs than baseline defenses operating in dense latent space. In particular, our method clearly outperforms dense mean-shift steering on all four models, and particularly against out-of-distribution attacks, showing that steering in sparse SAE feature space offers advantages over steering in dense activation space for jailbreak mitigation. Our results suggest off-the-shelf SAEs trained for interpretability can be repurposed as practical jailbreak defenses without task-specific training.

Hamid Bostani, Jacopo Cortellazzi, Daniel Arp et al.

Adversarial Training (AT) has been widely applied to harden learning-based classifiers against adversarial evasive attacks. However, its effectiveness in identifying and strengthening vulnerable areas of the model's decision space while maintaining high performance on clean data of malware classifiers remains an under-explored area. In this context, the robustness that AT achieves has often been assessed against unrealistic or weak adversarial attacks, which negatively affect performance on clean data and are arguably no longer threats. Previous work seems to suggest robustness is a task-dependent property of AT. We instead argue it is a more complex problem that requires exploring AT and the intertwined roles played by certain factors within data, feature representations, classifiers, and robust optimization settings, as well as proper evaluation factors, such as the realism of evasion attacks, to gain a true sense of AT's effectiveness. In our paper, we address this gap by systematically exploring the role such factors have in hardening malware classifiers through AT. Contrary to recent prior work, a key observation of our research and extensive experiments confirm the hypotheses that all such factors influence the actual effectiveness of AT, as demonstrated by the varying degrees of success from our empirical analysis. We identify five evaluation pitfalls that affect state-of-the-art studies and summarize our insights in ten takeaways to draw promising research directions toward better understanding the factors' settings under which adversarial training works at best.

Ilias Tsingenopoulos, Jacopo Cortellazzi, Branislav Bošanský et al.

ML-based malware detection on dynamic analysis reports is vulnerable to both evasion and spurious correlations. In this work, we investigate a specific ML architecture employed in the pipeline of a widely-known commercial antivirus company, with the goal to harden it against adversarial malware. Adversarial training, the sole defensive technique that can confer empirical robustness, is not applicable out of the box in this domain, for the principal reason that gradient-based perturbations rarely map back to feasible problem-space programs. We introduce a novel Reinforcement Learning approach for constructing adversarial examples, a constituent part of adversarially training a model against evasion. Our approach comes with multiple advantages. It performs modifications that are feasible in the problem-space, and only those; thus it circumvents the inverse mapping problem. It also makes possible to provide theoretical guarantees on the robustness of the model against a particular set of adversarial capabilities. Our empirical exploration validates our theoretical insights, where we can consistently reach 0% Attack Success Rate after a few adversarial retraining iterations.

Limin Yang, Zhi Chen, Jacopo Cortellazzi et al.

Malware classifiers are subject to training-time exploitation due to the need to regularly retrain using samples collected from the wild. Recent work has demonstrated the feasibility of backdoor attacks against malware classifiers, and yet the stealthiness of such attacks is not well understood. In this paper, we investigate this phenomenon under the clean-label setting (i.e., attackers do not have complete control over the training or labeling process). Empirically, we show that existing backdoor attacks in malware classifiers are still detectable by recent defenses such as MNTD. To improve stealthiness, we propose a new attack, Jigsaw Puzzle (JP), based on the key observation that malware authors have little to no incentive to protect any other authors' malware but their own. As such, Jigsaw Puzzle learns a trigger to complement the latent patterns of the malware author's samples, and activates the backdoor only when the trigger and the latent pattern are pieced together in a sample. We further focus on realizable triggers in the problem space (e.g., software code) using bytecode gadgets broadly harvested from benign software. Our evaluation confirms that Jigsaw Puzzle is effective as a backdoor, remains stealthy against state-of-the-art defenses, and is a threat in realistic settings that depart from reasoning about feature-space only attacks. We conclude by exploring promising approaches to improve backdoor defenses.

Jacopo Cortellazzi, Feargus Pendlebury, Daniel Arp et al.

Recent research efforts on adversarial machine learning (ML) have investigated problem-space attacks, focusing on the generation of real evasive objects in domains where, unlike images, there is no clear inverse mapping to the feature space (e.g., software). However, the design, comparison, and real-world implications of problem-space attacks remain underexplored. This article makes three major contributions. Firstly, we propose a general formalization for adversarial ML evasion attacks in the problem-space, which includes the definition of a comprehensive set of constraints on available transformations, preserved semantics, absent artifacts, and plausibility. We shed light on the relationship between feature space and problem space, and we introduce the concept of side-effect features as the by-product of the inverse feature-mapping problem. This enables us to define and prove necessary and sufficient conditions for the existence of problem-space attacks. Secondly, building on our general formalization, we propose a novel problem-space attack on Android malware that overcomes past limitations in terms of semantics and artifacts. We have tested our approach on a dataset with 150K Android apps from 2016 and 2018 which show the practical feasibility of evading a state-of-the-art malware classifier along with its hardened version. Thirdly, we explore the effectiveness of adversarial training as a possible approach to enforce robustness against adversarial samples, evaluating its effectiveness on the considered machine learning models under different scenarios. Our results demonstrate that "adversarial-malware as a service" is a realistic threat, as we automatically generate thousands of realistic and inconspicuous adversarial applications at scale, where on average it takes only a few minutes to generate an adversarial instance.