Deep neural networks (DNNs) represent critical intellectual property that model owners urgently need to protect. With the increasing value of models, malicious attackers increasingly attempt to extract model functionality through techniques such as fine-tuning, distillation, and pruning. Model fingerprinting has emerged as a mainstream protection strategy. However, existing fingerprinting methods either exhibit vulnerability to model modifications due to reliance on decision boundary features or require prohibitively large query budgets for accurate verification. This paper proposes ProbeSpec, which captures model fingerprints through dynamic behavioral analysis rather than static output matching. We discover that a model’s response patterns under multi-level perturbations form a unique “behavioral spectrum”, originating from implicit decision mechanisms learned during training and preserved even after various attacks. ProbeSpec employs three complementary probe types to elicit this characteristic and leverages DCT frequency-domain transformation for efficient fingerprint extraction. Extensive experiments show that ProbeSpec achieves 100% detection rate in the majority of attack scenarios, with an overall accuracy exceeding 95% across all tested architectures. Meanwhile, it effectively distinguishes independently trained models and requires only 80 probe samples for fingerprint extraction.
Deep neural networks are vulnerable to adversarial examples, even in hard-label black-box settings where only the top-1 prediction is available. To address the challenges of high-dimensional optimization under limited query budgets, we propose two query-efficient attack methods: Adam-OPT, which integrates Adam-based adaptive optimization into the ray-search framework to stabilize and accelerate zeroth-order gradient updates; Prior-Adam-OPT, which further incorporates transfer-based priors from surrogate models to enhance gradient estimation. Adam-OPT leverages historical gradient information and per-parameter adaptive updates to improve convergence, while Prior-Adam-OPT constructs a prior-guided orthogonal search basis that combines surrogate and random directions, enhancing both gradient accuracy and query efficiency. Our approach demonstrates superior performance across CIFAR-10, ImageNet, and zero-shot CLIP models, consistently reducing perturbation magnitudes and improving attack efficiency compared to state-of-the-art hard-label attacks. Ablation studies highlight the importance of the number of vectors used for gradient estimation and the quality of surrogate models, showing that combining adaptive optimization with transfer-based priors provides a scalable and robust framework for generating high-quality adversarial examples in challenging black-box scenarios.