On the Robustness of Diffusion-Based Image Compression to Bit-Flip Errors¶

Conference: CVPR 2026 arXiv: 2604.05743 Code: None (the paper mentions a reference implementation but provides no specific link) Area: Image Compression / Model Robustness Keywords: Diffusion Models, Image Compression, Bit-Flip Errors, Channel Robustness, Reverse Channel Coding

TL;DR¶

This paper presents the first systematic study of bit-flip robustness in diffusion-based image compression. It demonstrates that reverse channel coding (RCC)-based diffusion compression methods are inherently more resilient to bit-flip errors than traditional and learned codecs, and proposes Robust Turbo-DDCM, which independently encodes each atom index to further enhance robustness. At BER \(10^{-3}\), the proposed method maintains high reconstruction quality with only a marginal increase in BPP.

Background & Motivation¶

Background: Neural image compression has advanced significantly in recent years, achieving strong perceptual quality at very low bitrates. Diffusion models have emerged as a powerful paradigm for image compression, enabling state-of-the-art rate-distortion-perception trade-offs through end-to-end training, pretrained model reuse, or zero-shot approaches. Representative methods include DDCM, Turbo-DDCM, and DiffC, all RCC-based zero-shot diffusion compression frameworks.
Limitations of Prior Work: Real-world systems are susceptible to bit-flip errors (BFE) caused by transmission noise, hardware degradation, or even adversarial attacks (e.g., rowhammer). Even a small number of bit flips can severely degrade reconstruction quality or render files undecodable. Current practice relies on error-correcting codes (ECC), which increase the size of the compressed representation and worsen rate-distortion performance.
Key Challenge: Optimization of image compression methods has focused almost exclusively on the rate-distortion-perception trade-off, with robustness largely overlooked. Traditional codecs rely on variable-length entropy coding (e.g., Huffman, arithmetic coding), where a single bit error can cause decoding desynchronization and cascading error propagation across all subsequent symbols.
Goal: Can diffusion-based compression simultaneously offer higher compression ratios and stronger robustness? How can bit-flip robustness be further enhanced?
Key Insight: In RCC-based methods, the compressed representation encodes control signals that guide the denoising trajectory rather than pixel values or transform coefficients directly. This indirect representation may inherently tolerate small perturbations—a few bit flips may still produce similar guidance signals and reconstruction trajectories.
Core Idea: Replace the joint combinatorial (lexicographic) encoding in Turbo-DDCM with independent encoding of each atom index, so that a single bit flip affects at most one atom rather than the entire subset selection. This trades a marginal increase in BPP for a substantial gain in robustness.

Method¶

Overall Architecture¶

The framework is built upon zero-shot diffusion image compression via DDCM/Turbo-DDCM. The encoder selects codebook atoms at each denoising step to guide the reverse diffusion process toward the target image, and the sequence of atom indices constitutes the compressed bitstream. The decoder executes a deterministic reverse diffusion process using the shared codebook and received indices. The core contribution of this paper is an analysis of the sources of robustness and the proposal of a more robust encoding protocol.

Key Designs¶

DDCM Base Framework:
- Function: Foundation for zero-shot diffusion image compression.
- Mechanism: At each reverse diffusion step, \(K\) Gaussian noise vectors from a reproducible codebook \(\mathcal{C}_t\) replace stochastic sampling. During encoding, the atom most correlated with the denoising residual is selected: \(k_t = \arg\max_k \langle \mathbf{C}_t(k), \mathbf{x}_0 - \hat{\mathbf{x}}_{0|t} \rangle\). The index sequence forms the bitstream, with BPP = \(T\lceil\log_2 K\rceil\) / number of pixels.
- Design Motivation: Deterministic codebook selection replaces random noise sampling to enable information transmission—the encoder and decoder share the same codebook, and the indices constitute the entirety of the compressed information.
Vulnerability Analysis of Turbo-DDCM:
- Function: Identifies the robustness bottleneck in the Turbo-DDCM encoding protocol.
- Mechanism: Turbo-DDCM replaces single-atom selection with sparse approximation, selecting \(M\) atoms per step and encoding them as a single lexicographic index of \(\lceil\log_2\binom{K}{M}\rceil\) bits. A single bit flip can completely alter the decoded atom subset. For example, with \(K=8, M=3\), lexicographic index 0 corresponds to \(\{0,1,2\}\); flipping the most significant bit yields index 32, corresponding to \(\{1,4,7\}\)—a single bit error changes all three atoms.
- Design Motivation: Joint encoding is compression-efficient, but couples the information of multiple atoms, leading to error propagation effects.
Robust Turbo-DDCM:
- Function: Improves bit-flip robustness by independently encoding each atom index.
- Mechanism: Each selected atom index is independently encoded as \(\lceil\log_2 K\rceil\) bits instead of using joint lexicographic encoding. A single bit flip can then corrupt at most one atom's selection. BPP becomes \((T-1-N)(M\lceil\log_2 K\rceil + MC)\) / number of pixels, slightly higher than the original. Since the quality gain from increasing \(M\) exhibits diminishing returns, the loss from encoding fewer atoms at the same BPP is limited.
- Design Motivation: An explicit trade-off between quality and robustness—localizing error impact to exchange a minor compression efficiency loss for substantially improved error tolerance.

Loss & Training¶

This method is zero-shot and requires no training. It uses a pretrained Stable Diffusion 2.1 as the diffusion model. The compression and decompression pipeline is entirely based on deterministic codebook selection algorithms; the only modification is to the bitstream encoding protocol.

Key Experimental Results¶

Main Results¶

Reconstruction quality on Kodak24 at BER \(= 10^{-4}\):

Method	Type	BPP	PSNR (no noise)	PSNR (BER=1e-4)	File Corruption Rate
JPEG	Traditional	1.0	~30	Severe degradation	High
BPG	Traditional	0.5	~30	Severe degradation	High
ILLM	Learned	~0.1	~28	Severe degradation	High
StableCodec	Diffusion	~0.1	~25	Severe degradation	High
DDCM	RCC	~0.1	~24	Well preserved	0%
Turbo-DDCM	RCC	~0.1	~25	Slight degradation	0%
Robust T-DDCM	RCC	~0.1	~24	Near-lossless	0%

Ablation Study¶

Configuration	BER=1e-4 PSNR	BER=1e-3 PSNR	BER=1e-2 File Corruption Rate
JPEG	Severe degradation	Unusable	>80%
Turbo-DDCM	Slight degradation	Noticeable degradation	0%
Robust Turbo-DDCM	Near-lossless	Near-lossless	0%
Noise-free rate-distortion	Turbo-DDCM marginally better	—	—

Key Findings¶

Non-RCC methods exhibit sharp PSNR drops starting at BER ~\(10^{-5}\), whereas RCC methods degrade far more gradually.
Robust Turbo-DDCM maintains near-lossless reconstruction at BER \(= 10^{-3}\), a noise level at which all other methods have already severely degraded or become unusable.
On the file corruption rate metric, non-RCC methods exceed 80% corruption at BER ~\(10^{-2}\), while all RCC methods maintain 0% across the entire BER range.
The robustness advantage of RCC is not solely attributable to the absence of entropy coding—robustness differences are observable within groups of methods that both use and omit entropy coding.
Under noise-free conditions, Robust Turbo-DDCM achieves slightly lower rate-distortion-perception performance than Turbo-DDCM, which is the expected cost of trading compression efficiency for robustness.

Highlights & Insights¶

A "bonus" property of diffusion compression is discovered: RCC methods not only provide higher compression ratios but also exhibit inherently superior bit-flip robustness. This stems from the fact that the compressed representation encodes control signals for the denoising trajectory rather than raw data, so small perturbations may still produce similar trajectories.
The encoding protocol is critical to robustness: Modifying only the bitstream encoding scheme (joint → independent), without changing the model architecture or algorithmic logic, yields an order-of-magnitude improvement in robustness. This highlights that the importance of encoding protocol design in compression systems has been underestimated.
Potential to rethink the traditional compression–ECC separation pipeline: If the compressed representation is sufficiently robust intrinsically, weaker ECC or even no ECC may suffice, saving bandwidth and simplifying system design.

Limitations & Future Work¶

Only binary symmetric channel (BSC) independent bit flips are evaluated; burst errors and other structured channel models are not considered.
Some methods use entropy coding while DDCM/Turbo-DDCM do not, making it difficult to fully disentangle the contributions of representational robustness and the encoding scheme.
The encoding/decoding speed of RCC methods is far slower than traditional codecs (requiring full diffusion sampling), posing a practical barrier for real-time applications.
Evaluation is limited to Kodak24 and DIV2K; larger-scale or more diverse image datasets are not tested.
No comparison is made against joint source-channel coding (JSCC) methods.

vs. JPEG/BPG: Traditional codecs use variable-length entropy coding, where a single bit error can cause decoding desynchronization and cascading error propagation, resulting in very poor robustness.
vs. Turbo-DDCM: Robust Turbo-DDCM modifies only the encoding protocol, replacing joint lexicographic indexing with independent indexing, achieving near-lossless reconstruction at BER \(= 10^{-3}\) at the cost of a ~20% BPP increase.
vs. DiffC: DiffC, also an RCC method, similarly exhibits good robustness, but Robust Turbo-DDCM achieves further gains at high BER.
This work can inspire the wireless communications community to consider the inherent robustness of generative compression when designing end-to-end transmission systems.

Rating¶

Novelty: ⭐⭐⭐⭐ First systematic study of bit-flip robustness in diffusion-based compression, with findings that are both interesting and practically significant.
Experimental Thoroughness: ⭐⭐⭐⭐ Systematic evaluation across multiple BER values and compression method categories, though the datasets are limited.
Writing Quality: ⭐⭐⭐⭐⭐ Problem motivation is clearly articulated, analysis is accessible yet rigorous, and the explanation of the vulnerability mechanism (lexicographic encoding example) is highly intuitive.
Value: ⭐⭐⭐⭐ Reveals a new advantage dimension of diffusion-based compression with meaningful implications for communication and compression system design.