arXiv:2605.24550v1 Announce Type: new Abstract: Fine-tuning-as-a-Service (FaaS) enables personalization of large language models (LLMs), but it can weaken safety-alignment under harmful fine-tuning attacks. Recent work has shown that activating harmful-behavior modules during fine-tuning can prevent models from learning undesired behaviors, but its mechanism remains unclear. In this paper, we revisit temporary jailbreaking as a defense against harmful fine-tuning and provide a gradient-level analysis showing that it saturates safety-degrading gradients while preserving benign task-relevant gradients. Based on this insight, we propose a Buffer-and-Reinforce fine-tuning framework that buffers harmful updates during user fine-tuning and reinforces safety after adaptation. Specifically, BufferLoRA induces temporary jailbreaking as a removable adapter to reduce harmful updates during user fine-tuning. After adaptation, ReinforceLoRA, trained to recover refusal behavior under the temporarily jailbroken state, is integrated with UserLoRA via QR decomposition-based merging to reinforce safety while preserving user-task performance. Extensive experiments show that our framework achieves superior safety and utility with no additional safety data during user fine-tuning and minimal computational cost.

arXiv:2511.06325v2 Announce Type: replace-cross Abstract: Context-based detection methods such as DetectGPT achieve strong generalization in identifying AI-generated text by evaluating content compatibility with a model's learned distribution. In contrast, existing image detectors rely on discriminative features from pretrained backbones such as CLIP, which implicitly capture generator-specific artifacts. However, as modern generative models rapidly advance in visual fidelity, the artifacts these detectors depend on are becoming increasingly subtle or absent, undermining their reliability. Masked AutoEncoders (MAE) are inherently trained to reconstruct masked patches from visible context, naturally modeling patch-level contextual plausibility akin to conditional probability estimation, while also serving as a powerful semantic feature extractor through its encoder. We propose CINEMAE, a novel architecture that exploits both capabilities of MAE for AI-generated image detection: we derive per-patch anomaly signals from the reconstruction mechanism and extract global semantic features from the encoder, fusing both context-based and feature-based cues for robust detection. CINEMAE achieves highly competitive mean accuracies of 96.63\% on GenImage and 93.96\% on AIGCDetectBenchmark, maintaining over 93\% accuracy even under JPEG compression at QF=50.