Protecting the intellectual property of Large Language Models (LLMs) is critical because training them requires massive computational resources and data. A key challenge is determining whether a suspicious model is derived from a specific base model after fine-tuning or structural modification. Existing fingerprinting methods rely on model weights or high-dimensional representations, leading to substantial storage overhead. We propose a non-intrusive fingerprinting framework Concept REsponse Fingerprints (CREF). Inspired by activation engineering, CREF constructs a set of concept activation vectors as semantic probes. It then measures the response strength of hidden representations along these concept activation vectors using shared inputs. The resulting concept response matrix serves as a compact fingerprint, and similarity between models is measured using centered kernel alignment. Experiments on multiple LLM families show that CREF reliably distinguishes derived models from independently trained models and remains robust to fine-tuning, pruning, parameter permutation, and scaling. Moreover, the fingerprint requires only kilobyte-level storage, making it practical for large-scale deployment and ownership verification.

Chain-of-Thought (CoT) enhances an LLM’s ability to perform complex reasoning tasks, but it also introduces new security issues. In this work, we present ShadowCoT, a novel backdoor attack framework that targets the internal reasoning mechanism of LLMs. Unlike prior token-level or prompt-based attacks, ShadowCoT directly manipulates the model’s cognitive reasoning path, enabling it to hijack multi-step reasoning chains and produce logically coherent but adversarial outcomes. By conditioning on internal reasoning states, ShadowCoT learns to recognize and selectively disrupt key reasoning steps, effectively mounting a self-reflective cognitive attack within the target model. Our approach introduces a lightweight yet effective multi-stage injection pipeline, which selectively rewires attention pathways and perturbs intermediate representations with minimal parameter overhead (only 0.15% updated). ShadowCoT further leverages reinforcement learning and reasoning chain pollution (RCP) to autonomously synthesize stealthy adversarial CoTs that remain undetectable to advanced defenses. Extensive experiments across diverse reasoning benchmarks and LLMs show that ShadowCoT consistently achieves a state-of-the-art average Attack Success Rate of 91.2% (peaking at 94.4%) and a Hijacking Success Rate of 84.9% while preserving benign performance. These results reveal an emergent class of cognition-level threats and highlight the urgent need for defenses beyond shallowsurface-levelconsistency.

In this paper, we propose an end-to-end perceptual robust hashing scheme for video copy detection based on unsupervised learning. Firstly, the spatio-temporal information in videos is effectively fused and condensed into high-dimensional features through a 3D self-attention, multi-scale feature fusion model based on 3D-CNN, in which the Inception block and the 3D self-attention mechanism are integrated. Then, we calculate the correlation distances between the extracted features to differentiate perceptual contents. Based on the similarity relationship, we can dynamically generate the pseudo-labels and exploit them to further guide the model training for video hash generation. In addition, we design the dual constraints to make the hash code obtain satisfactory robustness and discrimination. Extensive experiments demonstrate that the proposed scheme achieves superior performance of copy detection compared with existing schemes and performs well even in the case of untrained manipulations.