Research map / Adversarial robustness and security / Certified and empirical robustness
Robust Image Hashing Based on Contrastive Masked Autoencoder with Weak-Strong Augmentation Alignment
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CMAA, an end-to-end contrastive masked autoencoder (ViT-Small with masking and a parameter-free binary quantization layer) with weak-strong augmentation alignment via KL distribution divergence for robust image hashing.
Problem
Robust image hashing for content identification must stay discriminative while resisting large-scale and hybrid attacks, and existing methods lose information in binarization and struggle with strong or hybrid attacks.
Approach
Combines a contrastive loss with a KL-based weak-strong augmentation alignment that makes strongly attacked views inherit the discrimination of weakly attacked views, plus a masked ViT backbone and an end-to-end differentiable binary quantization layer trained with a quantization loss.
Key results
- Average F1 of 0.9946 on CASIA and 0.9982 on Copydays, versus 0.9042 for the best baseline (Swin-T) on CASIA.
- Average NHDr robustness of 0.0987 on CASIA and 0.0905 on Copydays, lower (better) than all five baselines.
- Ablation: removing LD and LR, or LQ, lowers F1 on CASIA from 0.9946 to 0.9641 and 0.9918 respectively.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LD and LR (compared with omitting both) (ablation) | CMAA · CASIA | NHDd average (µ): 0.4962 -> 0.4947 | No clear effect |
| add LD and LR (compared with omitting both) (ablation) | CMAA · CASIA | NHDr: 0.1343 -> 0.0987 | Helped |
| add LD and LR (compared with omitting both) (ablation) | CMAA · CASIA | F1 score: 0.9641 -> 0.9946 | Helped |
| add LR (ablation) | CMAA · CASIA | NHDd average (µ): 0.4905 -> 0.4947 | Helped |
| add LR (ablation) | CMAA · CASIA | NHDr: 0.1060 -> 0.0987 | Helped |
| add LR (ablation) | CMAA · CASIA | F1 score: 0.9906 -> 0.9946 | Helped |
| add LD (ablation) | CMAA · CASIA | NHDd average (µ): 0.4930 -> 0.4947 | Helped |
| add LD (ablation) | CMAA · CASIA | NHDr: 0.1122 -> 0.0987 | Helped |
| add LD (ablation) | CMAA · CASIA | F1 score: 0.9856 -> 0.9946 | Helped |
| add LQ (ablation) | CMAA · CASIA | NHDd average (µ): 0.4917 -> 0.4947 | Helped |
| add LQ (ablation) | CMAA · CASIA | NHDr: 0.0946 -> 0.0987 | Hurt |
| add LQ (ablation) | CMAA · CASIA | F1 score: 0.9918 -> 0.9946 | Helped |
Surprising findings
- Adding the KL alignment and reconstruction losses without LQ yields slightly different NHDd, and the paper notes equal loss weights performed well without tuning.
Limitations
- Discrimination (NHDd mean) is only comparable to ISM and is slightly below ISM on Copydays
- Hyperparameter sensitivity and time complexity are deferred to supplementary material
- Evaluated only on CASIA and Copydays in the main text; UCID is in supplementary
Experimental setup
Datasets: COCO mini-train
Benchmarks: CASIA Copydays
Metrics: NHDd NHDr F1 score TPR/FPR DET curves
Baselines: ProtoHash TMICH Swin-T GLIF ISM
Builds on: MoCo MAE SimSiam ProtoHash Li and van Gemert binary layer
Closest papers by meaning
- An Asymmetric Augmented Self-Supervised Learning Method for Unsupervised Fine-Grained Image Hashing (CVPR 2024 · similarity 0.84)
- Decoupled Prototype Contrastive Alignment Hashing for Cross-Modal Retrieval (NeurIPS 2026 · similarity 0.84)
- HUANG: A Robust Diffusion Model-based Targeted Adversarial Attack Against Deep Hashing Retrieval (AAAI 2025 · similarity 0.83)
- Dual Self-Paced Cross-Modal Hashing (AAAI 2024 · similarity 0.82)
- Contrastive Tuning: A Little Help to Make Masked Autoencoders Forget (AAAI 2024 · similarity 0.82)
- A Lottery Ticket Hypothesis Approach with Sparse Fine-tuning and MAE for Image Forgery Detection and Localization (AAAI 2025 · similarity 0.82)
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.