Bitstream Collisions in Neural Image Compression via Adversarial Perturbations
Not on the map. Why?
Abstract (arXiv)
Neural image compression (NIC) has emerged as a promising alternative to classical compression techniques, offering improved compression ratios. Despite its progress towards standardization and practical deployment, there has been minimal exploration into it's robustness and security. This study reveals an unexpected vulnerability in NIC - bitstream collisions - where semantically different images produce identical compressed bitstreams. Utilizing a novel whitebox adversarial attack algorithm, this paper demonstrates that adding carefully crafted perturbations to semantically different images can cause their compressed bitstreams to collide exactly. The collision vulnerability poses a threat to the practical usability of NIC, particularly in security-critical applications. The cause of the collision is analyzed, and a simple yet effective mitigation method is presented.
Key facts. Cited by 1 paper in the map (data through 14 Oct 2026). Builds on 33 of its 42 references that are in the map. First posted on arXiv 25 Mar 2025.
Builds on
33 of 42 references.
References · 33
- 2017Towards Deep Learning Models Resistant to Adversarial Attackscited by 6,415
- 2013Intriguing properties of neural networkscited by 5,155
- 2016SGDR: Stochastic Gradient Descent with Warm Restartscited by 5,045
- 2016Towards Evaluating the Robustness of Neural Networkscited by 2,998
- 2018Variational image compression with a scale hyperpriorcited by 935
- 2016End-to-end Optimized Image Compressioncited by 696
- 2018Joint Autoregressive and Hierarchical Priors for Learned Image Compressioncited by 664
- 2020Learned Image Compression with Discretized Gaussian Mixture Likelihoods and Attention Modulescited by 590
- 2018Black-box Adversarial Attacks with Limited Queries and Informationcited by 501
- 2017Decision-Based Adversarial Attacks: Reliable Attacks Against Black-Box Machine Learning Modelscited by 462
- 2020High-Fidelity Generative Image Compressioncited by 338
- 2019HopSkipJumpAttack: A Query-Efficient Decision-Based Attackcited by 333
21 more
- 2017Lossy Image Compression with Compressive Autoencoderscited by 294
- 2019Efficient and Accurate Estimation of Lipschitz Constants for Deep Neural Networkscited by 293
- 2016Full Resolution Image Compression with Recurrent Neural Networkscited by 175
- 2015Variable Rate Image Compression with Recurrent Neural Networkscited by 173
- 2021Implicit Neural Representations for Image Compressioncited by 162
- 2017Soft-to-Hard Vector Quantization for End-to-End Learning Compressible Representationscited by 152
- 2017Real-Time Adaptive Image Compressioncited by 130
- 2015Density Modeling of Images using a Generalized Normalization Transformationcited by 127
- 2023Improving Statistical Fidelity for Neural Image Compression with Implicit Local Likelihood Modelscited by 94
- 2019Image and Video Compression with Neural Networks: A Reviewcited by 86
- 2020Improving Inference for Neural Image Compressioncited by 83
- 2022An Introduction to Neural Data Compressioncited by 79
- 2016End-to-end optimization of nonlinear transform codes for perceptual qualitycited by 73
- 2021Soft then Hard: Rethinking the Quantization in Neural Image Compressioncited by 51
- 2022Unified Multivariate Gaussian Mixture for Efficient Neural Image Compressioncited by 43
- 2017Generative Compressioncited by 40
- 2021Slimmable Compressive Autoencoders for Practical Neural Image Compressioncited by 39
- 2021Towards Robust Neural Image Compression: Adversarial Attack and Model Finetuningcited by 25
- 2022Flexible Neural Image Compression via Code Editingcited by 20
- 2022Learning-Driven Lossy Image Compression; A Comprehensive Surveycited by 13
- 2024Exploring adversarial robustness of JPEG AI: methodology, comparison and new methodscited by 3
Built on by
Citing papers per year: 2026: 1.
1 paper.