{"paper":{"title":"DGS-Net: Distillation-Guided Gradient Surgery for CLIP Fine-Tuning in AI-Generated Image Detection","license":"http://creativecommons.org/licenses/by/4.0/","headline":"By projecting task gradients onto the orthogonal complement of harmful directions and aligning with beneficial ones distilled from a frozen CLIP encoder, DGS-Net fine-tunes for AI-generated image detection without catastrophic forgetting.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Boyu Wang, Fan Wang, Jiazhen Yan, Zhangjie Fu, Ziqiang Li, Ziwen He","submitted_at":"2025-11-17T08:05:31Z","abstract_excerpt":"The rapid progress of generative models such as GANs and diffusion models has led to the widespread proliferation of AI-generated images, raising concerns about misinformation, privacy violations, and trust erosion in digital media. Although large-scale multimodal models like CLIP offer strong transferable representations for detecting synthetic content, fine-tuning them often induces catastrophic forgetting, which degrades pre-trained priors and limits cross-domain generalization. To address this issue, we propose the Distillation-guided Gradient Surgery Network (DGS-Net), a novel framework t"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"By projecting task gradients onto the orthogonal complement of harmful directions and aligning with beneficial ones distilled from a frozen CLIP encoder, DGS-Net achieves unified optimization of prior preservation and irrelevant suppression, outperforming state-of-the-art approaches by an average margin of 6.6% across 50 generative models.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the proposed gradient-space decomposition can reliably and consistently separate harmful from beneficial descent directions without introducing instability or requiring extensive hyperparameter tuning specific to each dataset.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"DGS-Net applies distillation-guided gradient surgery during CLIP fine-tuning to preserve pre-trained knowledge and suppress irrelevant features, reporting 6.6% average gains over prior methods on 50 generative models.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"By projecting task gradients onto the orthogonal complement of harmful directions and aligning with beneficial ones distilled from a frozen CLIP encoder, DGS-Net fine-tunes for AI-generated image detection without catastrophic forgetting.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"3f9963f3ae5b797f1a281820499238d5cbbdb886f0f27461d9399c1327e5f523"},"source":{"id":"2511.13108","kind":"arxiv","version":4},"verdict":{"id":"b3fb5e10-f9a6-4f4d-b731-2806db10627e","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T22:07:21.866370Z","strongest_claim":"By projecting task gradients onto the orthogonal complement of harmful directions and aligning with beneficial ones distilled from a frozen CLIP encoder, DGS-Net achieves unified optimization of prior preservation and irrelevant suppression, outperforming state-of-the-art approaches by an average margin of 6.6% across 50 generative models.","one_line_summary":"DGS-Net applies distillation-guided gradient surgery during CLIP fine-tuning to preserve pre-trained knowledge and suppress irrelevant features, reporting 6.6% average gains over prior methods on 50 generative models.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the proposed gradient-space decomposition can reliably and consistently separate harmful from beneficial descent directions without introducing instability or requiring extensive hyperparameter tuning specific to each dataset.","pith_extraction_headline":"By projecting task gradients onto the orthogonal complement of harmful directions and aligning with beneficial ones distilled from a frozen CLIP encoder, DGS-Net fine-tunes for AI-generated image detection without catastrophic forgetting."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2511.13108/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"01533e939b13e000cb3846fbaceea91b21b43ccd8f784f1e9a16b82fc82f1d39"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}