Kehuan Zhang

Android’s vast ecosystem and expansive API surfaces pose a serious challenge to static application security testing (SAST) tools. Mainstream tools such as MobSF, APKHunt, and AUSERA mainly rely on manually crafted rules. Crafting these rules demands considerable effort, yet they still cannot cover every security issue. When Android introduces new APIs, changes its permission model, or revises other security policies, the rules soon fall behind. Without constant maintenance, false positives grow, and true vulnerabilities slip through. Recently released LLM-based detectors are easy to use and potentially support a wide range of vulnerability types, but their findings often lack clear explanations and suffer from high false-positive rates. In this paper, we present RULEDROID, a new framework that leverages LLMs to automatically generate Semgrep-compatible static detection rules from up-to-date official Android security documentation. RULEDROID tackles the limits of pure LLM detection by combining (i) Retrieval-Augmented Generation (RAG), which grounds model outputs in trusted documents, and (ii) a modular workflow that decomposes rule synthesis into welldefined stages. The resulting rules are then applied with proven static-analysis techniques, ensuring consistent and explainable results. We evaluated RULEDROID on three public benchmark datasets. Based on its large and precise rule set, RULEDROID achieved higher coverage and accuracy than traditional SAST tools, and sharply reduced false positives compared with direct LLM scanning. When applied to real-world apps, RULEDROID discovered multiple new vulnerabilities, resulting in 57 CVE IDs being assigned. These results show that RULEDROID combines the broad vulnerability coverage of LLMs with the precision of static analysis, delivering a fully automated docs-to-rules solution for Android security testing.

Deep Neural Networks remain inherently vulnerable to backdoor attacks. Traditional test-time defenses largely operate under the paradigm of internal diagnosis methods like model repairing or input robustness, yet these approaches are often fragile under advanced attacks as they remain entangled with the victim model's corrupted parameters. We propose a paradigm shift from Internal Diagnosis to External Semantic Auditing, arguing that effective defense requires decoupling safety from the victim model via an independent, semantically grounded auditor. To this end, we present a framework harnessing Universal Vision-Language Models (VLMs) as evolving semantic gatekeepers. We introduce PRISM (Prototype Refinement & Inspection via Statistical Monitoring), which overcomes the domain gap of general VLMs through two key mechanisms: a Hybrid VLM Teacher that dynamically refines visual prototypes online, and an Adaptive Router powered by statistical margin monitoring to calibrate gating thresholds in real-time. Extensive evaluation across 17 datasets and 11 attack types demonstrates that PRISM achieves state-of-the-art performance, suppressing Attack Success Rate to <1% on CIFAR-10 while improving clean accuracy, establishing a new standard for model-agnostic, externalized security.

Clean-image backdoor attacks, which use only label manipulation in training datasets to compromise deep neural networks, pose a significant threat to security-critical applications. A critical flaw in existing methods is that the poison rate required for a successful attack induces a proportional, and thus noticeable, drop in Clean Accuracy (CA), undermining their stealthiness. This paper presents a new paradigm for clean-image attacks that minimizes this accuracy degradation by optimizing the trigger itself. We introduce Generative Clean-Image Backdoors (GCB), a framework that uses a conditional InfoGAN to identify naturally occurring image features that can serve as potent and stealthy triggers. By ensuring these triggers are easily separable from benign task-related features, GCB enables a victim model to learn the backdoor from an extremely small set of poisoned examples, resulting in a CA drop of less than 1%. Our experiments demonstrate GCB's remarkable versatility, successfully adapting to six datasets, five architectures, and four tasks, including the first demonstration of clean-image backdoors in regression and segmentation. GCB also exhibits resilience against most of the existing backdoor defenses.

Reconstruction attackers can exploit facial features to recover the original user’s face, resulting in user privacy leakage. One new strategy to enhance the “Edge-Cloud” face recognition system’s privacy is to add adversarial perturbations to facial features, preventing the attackers from high-quality user image recovery. However, the existing works following this strategy suffer from unacceptable damage to face recognition accuracy. Achieving robust privacy enhancement and face recognition accuracy simultaneously is still challenging. To tackle this challenge, we propose an adversarial perturbation-based plug-and-play privacy-enhancing method (Patronus) with robustness against face image reconstruction attacks and near-lossless face recognition performance. The key insight is derived from our observation that the feature distance between two face images of the same person is significantly lower than the threshold set in the face recognition system. This leaves room for adding adversarial perturbations to the facial features without compromising face recognition accuracy. Our strategy limits the amount of adversarial perturbations in a fine-grained manner to ensure that they are within the range of not damaging face recognition accuracy. Our evaluation shows the superior performance of Patronus in robustness against reconstruction attacks and near-lossless face recognition accuracy compared to state-of-the-art (SOTA) methods. Patronus can be easily integrated into deployed face recognition systems as a plug-in privacy-enhancing module with low overhead.

Analyzing binary code is vital for software engineering and security research, particularly when the source code is unavailable. However, understanding, modifying, and retargeting binary code can be complex tasks. To counter these difficulties, binary lifters have been introduced. These tools translate binary code into Intermediate Representations (IRs), providing several advantages, such as enabling modifications to executables without source code and facilitating code retargetability. So far, accurately developing binary lifters for modern ISAs is universally acknowledged as challenging and error-prone. Existing validation methods mainly concentrate on isolated instructions, overlooking interactions among instructions. In this paper, we introduce LiftFuzz, a novel framework that leverages instruction context-aware fuzzing to validate binary lifters. LiftFuzz harnesses an assembly language model to learn interactions among instructions and generates test cases with the knowledge. LiftFuzz greatly outperforms the baseline, requiring only 1/1000 of the test cases used by the baseline to identify 26 inconsistencies, including a previously uncovered category. LiftFuzz significantly contributes to enhancing the performance of binary lifters, which are frequently employed in binary security applications.