Pyo Min Hong

We propose dgMARK, a decoding-guided watermarking method for discrete diffusion language models (dLLMs). Unlike autoregressive models, dLLMs can generate tokens in arbitrary order. While an ideal conditional predictor would be invariant to this order, practical dLLMs exhibit strong sensitivity to the unmasking order, creating a new channel for watermarking. dgMARK steers the unmasking order toward positions whose high-reward candidate tokens satisfy a simple parity constraint induced by a binary hash, without explicitly reweighting the model’s learned probabilities. The method is plug-and-play with common decoding strategies (e.g., confidence, entropy, and margin-based ordering) and can be strengthened with a one-step lookahead variant. Watermarks are detected via elevated parity-matching statistics, and a sliding-window detector ensures robustness under post-editing operations including insertion, deletion, substitution, and paraphrasing.

Recent threats to deep learning-based biometric authentication systems stem from adversarial attacks exploiting vulnerabilities in deep learning models. While existing studies extensively analyze the risk of such attacks, they primarily focus on isolated modules (e.g., liveness detectors or identity matchers) or specific adversarial attack types (e.g., evasion and poisoning attacks). In this paper, we introduce a novel approach that comprehensively assesses the risk of adversarial attacks by simulating multiple scenarios within biometric authentication systems. We identify the surfaces susceptible to adversarial attacks within these systems and devise scenarios that reflect the dependencies between modules. Moreover, we establish evaluation metrics to comprehensively assess the risk involved. Through a case study conducted on a real-world face recognition system, we successfully demonstrate the effectiveness of our approach. Our approach facilitates the systematic evaluation of the security of target biometric authentication systems against adversarial attacks. Ultimately, it enables the establishment of robust and proactive defense mechanisms.

The development and deployment of deep learning models are currently limited to expert developers due to the requirement for programming expertise, posing challenges for non-expert or inexperienced developers. In this paper, we present VELCRO, a visual-based programming tool that enables intuitive construction, modification, analysis, and generation of deep learning models for both expert and non-expert developers. VELCRO allows users to understand and modify complex architectures of deep learning models using architecture abstraction, and verify architectures through automated architecture validation. Finally, users can convert the user-configured architecture into deep learning code and deploy it in target environments.