In this paper, we introduce GradEscape, the first gradient-based evader designed to attack AI-generated text (AIGT) detectors. GradEscape overcomes the undifferentiable computation problem, caused by the discrete nature of text, by introducing a novel approach to construct weighted embeddings for the detector input. It then updates the evader model parameters using feedback from victim detectors, achieving high attack success with minimal text modification. To address the issue of tokenizer mismatch between the evader and the detector, we introduce a warm-started evader method, enabling GradEscape to adapt to detectors across any language model architecture. Moreover, we employ novel tokenizer inference and model extraction techniques, facilitating effective evasion even in query-only access. We evaluate GradEscape on four datasets and three widely-used language models, benchmarking it against four state-of-the-art AIGT evaders. Experimental results demonstrate that GradEscape outperforms existing evaders in various scenarios, including with an 11B paraphrase model, while utilizing only 139M parameters. We have successfully applied GradEscape to two real-world commercial AIGT detectors. Our analysis reveals that the primary vulnerability stems from disparity in text expression styles within the training data. We also propose a potential defense strategy to mitigate the threat of AIGT evaders. We open-source our GradEscape for developing more robust AIGT detectors.

Latent-space monitors aim to detect undesirable behaviours in Large Language Models by leveraging their internal representations rather than relying solely on black-box outputs. These methods have shown promise in identifying behaviours such as deception and unsafe completions. However, these monitors may themselves become training signals, for example, by using problematic samples found in deployment to retrain models. This raises an important question: can models learn to evade such monitors? To evaluate this capability, we introduce RL-Obfuscation, in which LLMs are finetuned via reinforcement learning to evade latent-space monitors while maintaining their blackbox behaviour. We apply RL-Obfuscation to Language Models ranging from 7B to 14B parameters and evaluate their Evasion Success Rate against a suite of monitors. We find that token-level monitors are highly vulnerable to this attack while more holistic monitors, such as max-pooling or attention-based probes, remain robust. Moreover, for these vulnerable monitors, models trained to evade a single static monitor can generalise to evade other unseen monitors. We also find that the models can be trained to conditionally bypass latent-space monitors on only certain inputs. Finally, we study how the models bypass these monitors and find that the model can learn to repurpose tokens to have different internal representations.

Large Language Models (LLMs) can be \emph{misused} to spread online spam and misinformation. Content watermarking deters misuse by hiding a message in model-generated outputs, enabling their detection using a secret watermarking key. Robustness is a core security property, stating that evading detection requires (significant) degradation of the content's quality. Many LLM watermarking methods have been proposed, but robustness is tested only against \emph{non-adaptive} attackers who lack knowledge of the watermarking method and can find only suboptimal attacks. We formulate the robustness of LLM watermarking as an objective function and propose preference-based optimization to tune \emph{adaptive} attacks against the specific watermarking method. Our evaluation shows that (i) adaptive attacks substantially outperform non-adaptive baselines. (ii) Even in a non-adaptive setting, adaptive attacks optimized against a few known watermarks remain highly effective when tested against other unseen watermarks, and (iii) optimization-based attacks are practical and require less than seven GPU hours. Our findings underscore the need to test robustness against adaptive attackers.

Machine learning malware detectors are vulnerable to adversarial EXEmples, i.e. carefully-crafted Windows programs tailored to evade detection. Unlike other adversarial problems, attacks in this context must be functionality-preserving, a constraint which is challenging to address. As a consequence heuristic algorithms are typically used, that inject new content, either randomly-picked or harvested from legitimate programs. In this paper, we show how learning malware detectors can be cast within a zeroth-order optimization framework which allows to incorporate functionality-preserving manipulations. This permits the deployment of sound and efficient gradient-free optimization algorithms, which come with theoretical guarantees and allow for minimal hyper-parameters tuning. As a by-product, we propose and study ZEXE, a novel zero-order attack against Windows malware detection. Compared to state-of-the-art techniques, ZEXE provides drastic improvement in the evasion rate, while reducing to less than one third the size of the injected content.

Generative AI (GenAI) can synthesize extremely realistic-looking images, posing growing challenges to information authenticity on the Internet. Watermarking [1-7] was suggested as a key technology to distinguish AI-generated and non-AI-generated content in the Executive Order on AI security issued by the White House in October 2023. In watermarkbased detection, a watermark is embedded into an AI-generated image before releasing it; and an image is detected as AI-generated if the same watermark can be decoded from it. Watermarking AI-generated images has been widely deployed in industry. For instance, Google's SynthID watermarks images generated by Imagen [8]; OpenAI embeds a watermark into images generated by DALL-E [9]; and Stable Diffusion enables users to embed a watermark into the generated images [10]. An attacker can use evasion attacks [11] to remove the watermark in a watermarked image to evade detection. Specifically, an evasion attack strategically adds a perturbation into a watermarked image such that the target watermark-based detector falsely detects the perturbed image as non-AI-generated. The literature has well understood the robustness of watermark-based detector against evasion attacks in the white-box setting (i.e., the attacker has access to the target watermarking model) and black-box setting (i.e., the attacker has access to the detection API) [11]. Specifically, in the white-box setting, an attacker can find a small perturbation for a given watermarked image such that the perturbed image evades detection while maintaining the image's visual quality; and in the

Adversarial Malware Generation (AMG), the generation of adversarial malware variants to strengthen Deep Learning (DL)-based malware detectors has emerged as a crucial tool in the development of proactive cyberdefense. However, the majority of extant works offer subtle perturbations or additions to executable files and do not explore full-file obfuscation. In this study, we show that an open-source encryption tool coupled with a Reinforcement Learning (RL) framework can successfully obfuscate malware to evade state-of-the-art malware detection engines and outperform techniques that use advanced modification methods. Our results show that the proposed method improves the evasion rate from 27%-49% compared to widely-used state-of-the-art reinforcement learning-based methods.

Due to the proliferation of malware, defenders are increasingly turning to automation and machine learning as part of the malware detection tool-chain. However, machine learning models are susceptible to adversarial attacks, requiring the testing of model and product robustness. Meanwhile, attackers also seek to automate malware generation and evasion of antivirus systems, and defenders try to gain insight into their methods. This work proposes a new algorithm that combines Malware Evasion and Model Extraction (MEME) attacks. MEME uses model-based reinforcement learning to adversarially modify Windows executable binary samples while simultaneously training a surrogate model with a high agreement with the target model to evade. To evaluate this method, we compare it with two state-of-the-art attacks in adversarial malware creation, using three well-known published models and one antivirus product as targets. Results show that MEME outperforms the state-of-the-art methods in terms of evasion capabilities in almost all cases, producing evasive malware with an evasion rate in the range of 32-73%. It also produces surrogate models with a prediction label agreement with the respective target models between 97-99%. The surrogate could be used to fine-tune and improve the evasion rate in the future.

Machine learning has proven to be a useful tool for automated malware detection, but machine learning models have also been shown to be vulnerable to adversarial attacks. This article addresses the problem of generating adversarial malware samples, specifically malicious Windows Portable Executable files. We summarize and compare work that has focused on adversarial machine learning for malware detection. We use gradient-based, evolutionary algorithm-based, and reinforcement-based methods to generate adversarial samples, and then test the generated samples against selected antivirus products. We compare the selected methods in terms of accuracy and practical applicability. The results show that applying optimized modifications to previously detected malware can lead to incorrect classification of the file as benign. It is also known that generated malware samples can be successfully used against detection models other than those used to generate them and that using combinations of generators can create new samples that evade detection. Experiments show that the Gym-malware generator, which uses a reinforcement learning approach, has the greatest practical potential. This generator achieved an average sample generation time of 5.73 seconds and the highest average evasion rate of 44.11%. Using the Gym-malware generator in combination with itself improved the evasion rate to 58.35%.

Deep learning (DL) methods have been widely applied to anomaly-based network intrusion detection system (NIDS) to detect malicious traffic. To expand the usage scenarios of DL-based methods, federated learning (FL) allows multiple users to train a global model on the basis of respecting individual data privacy. However, it has not yet been systematically evaluated how robust FL-based NIDSs are against existing privacy attacks under existing defenses. To address this issue, we propose two privacy evaluation metrics designed for FL-based NIDSs, including (1) privacy score that evaluates the similarity between the original and recovered traffic features using reconstruction attacks, and (2) evasion rate against NIDSs using adversarial attack with the recovered traffic. We conduct experiments to illustrate that existing defenses provide little protection and the corresponding adversarial traffic can even evade the SOTA NIDS Kitsune. To defend against such attacks and build a more robust FL-based NIDS, we further propose FedDef, a novel optimization-based input perturbation defense strategy with theoretical guarantee. It achieves both high utility by minimizing the gradient distance and strong privacy protection by maximizing the input distance. We experimentally evaluate four existing defenses on four datasets and show that our defense outperforms all the baselines in terms of privacy protection with up to 7 times higher privacy score, while maintaining model accuracy loss within 3% under optimal parameter combination.

ML models are known to be vulnerable to adversarial query attacks. In these attacks, queries are iteratively perturbed towards a particular class without any knowledge of the target model besides its output. The prevalence of remotely-hosted ML classification models and Machine-Learning-as-a-Service platforms means that query attacks pose a real threat to the security of these systems. To deal with this, stateful defenses have been proposed to detect query attacks and prevent the generation of adversarial examples by monitoring and analyzing the sequence of queries received by the system. Several stateful defenses have been proposed in recent years. However, these defenses rely solely on similarity or out-of-distribution detection methods that may be effective in other domains. In the malware detection domain, the methods to generate adversarial examples are inherently different, and therefore we find that such detection mechanisms are significantly less effective. Hence, in this paper, we present MalProtect, which is a stateful defense against query attacks in the malware detection domain. MalProtect uses several threat indicators to detect attacks. Our results show that it reduces the evasion rate of adversarial query attacks by 80+\% in Android and Windows malware, across a range of attacker scenarios. In the first evaluation of its kind, we show that MalProtect outperforms prior stateful defenses, especially under the peak adversarial threat.