The growing realism and accessibility of manipulated and generated faces threaten the trustworthiness of digital media. To detect such forgeries, deepfake detectors based on vision foundation models have shown promising performance, but they typically rely on a single pretrained representation and are prone to overfitting to particular training distributions. To improve generalization to unseen forgeries, we propose UCF-Net, an uncertainty-aware cascaded fusion network that harnesses CLIP's language-aligned semantic priors and DINO's self-supervised visual-structure priors. UCF-Net extracts hierarchical features across Transformer depths, uses layer-wise expert aggregation to adaptively combine each encoder's multi-level cues, and performs weighted fusion of the resulting representations based on entropy-derived uncertainty. We further consolidate public deepfake datasets into a unified benchmark of approximately 4M images and construct a separate cross-generator evaluation set with over 8K face images from eight recent generators. On the unified benchmark, UCF-Net achieves the best mean AUC among the evaluated methods in both in-domain and cross-domain evaluations. On the cross-generator set, it adapts effectively with limited target-domain data, although zero-shot transfer remains challenging.
Malicious URL detection serves as a critical component in safeguarding cyberspace security. Although deep learning models have achieved remarkable performance in identifying complex malicious patterns, their large parameter size and high computational overhead have become bottlenecks restricting their real-time deployment in resource-constrained environments such as gateways and mobile terminals. To address this challenge, this paper proposes Purl-Distill, a lightweight malicious URL detection framework based on multi-granularity knowledge distillation. Different from traditional compression methods that only focus on output alignment, Purl-Distill introduces a multi-granularity alignment mechanism. First, by constructing a multi-layer feature mapping space, it deeply transfers the hierarchical representations of the teacher model from low-level character morphology to high-level semantic logic. Second, a temperature-scaled logits distribution alignment strategy and a joint loss function are designed to accurately capture the decision boundary features of the teacher model. Experimental results demonstrate that on multiple public benchmark datasets, Purl-Distill achieves an order-of-magnitude reduction in model parameters and storage space while maintaining detection accuracy comparable to state-of-the-art (SOTA) teacher models. This work provides an efficient engineering solution for high-performance network security protection in resource-constrained scenarios.
