Research map / Vision-language and multimodal / Multimodal LLMs
TrustCLIP: Learning Private Visual Features via Adversarial Reconstruction
Nikos Athanasiou, Ilya A. Petrov, Angela Yao, Shugao Ma, Eric Sauser, Edoardo Remelli, Shreyas Hampali, Johannes Schönberger, Fadime Sener, Bugra Tekin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add adversarially trained privacy projection to CLIP features | TrustCLIP · SUN397 | DreamSim (privacy): 0.215 -> 0.514 | Improved |
| add adversarially trained privacy projection to CLIP features | TrustCLIP · SUN397 | Top-1 accuracy: 83.36 -> 82.915 | No clear effect |
| use learned projection instead of Gaussian noise | TrustCLIP · SUN397 | privacy-utility: DSIM 0.68 at 17.4% Top-1 (noise 1.0) vs 0.88 at 85.1% | Improved |
| add Gaussian noise sigma up to 0.1 (ablation) | CLIP features · SUN397 | DreamSim: 0.34 -> 0.33 | No clear effect |
| use standard MLP projection instead of identity-initialized (ablation) | TrustLLaVA · MM-Vet | score: 34.0 -> 27.8 | Hurt |
| use standard MLP projection instead of identity-initialized (ablation) | TrustLLaVA · VLM reconstruction attack | DreamSim: 0.43 -> 0.62 | Improved |
| add TrustCLIP projection to LLaVA-SP | TrustLLaVA · MME-Perception | score: 1470.7 -> 1390.8 | Hurt |
| add TrustCLIP projection to LLaVA-SP | TrustLLaVA · LLaVA-Bench | score: 61.7 -> 65.3 | Improved |
| add TrustCLIP projection to LLaVA-SP | TrustLLaVA · VQAv2 | score: 79.2 -> 76.3 | Hurt |
| add TrustCLIP projection to LLaVA-SP | TrustLLaVA · VLM reconstruction attack | DreamSim: 0.32 -> 0.39 | Improved |
| vary lambda_rec from 0.1 to 0.001 (ablation) | TrustLLaVA · VLM benchmarks | MM-Vet: 33.1 -> 34.0 | No clear effect |
| raise lambda_rec on SUN397 from 0.25 to 1 with L2+LPIPS (ablation) | TrustCLIP · SUN397 | DreamSim: 0.514 -> 0.799 | Improved |
Closest papers by meaning
- MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models (ICML 2026 · similarity 0.86)
- MultiTrust: A Comprehensive Benchmark Towards Trustworthy Multimodal Large Language Models (NeurIPS 2024 · similarity 0.85)
- Steal the Patch Size: Adversarially Manipulate Vision Language Models (ICML 2026 · similarity 0.85)
- ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time (ICLR 2025 · similarity 0.85)
- Private Attribute Inference from Images with Vision-Language Models (NeurIPS 2024 · similarity 0.85)
- VMDT: Decoding the Trustworthiness of Video Foundation Models (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
