TL;DR
This paper investigates the surprising vulnerability of deep learning-based image compression models to adversarial attacks that drastically increase file size (bitrate) with imperceptible input changes. The authors propose novel white-box and black-box attack methods, demonstrating significant bitrate inflation on state-of-the-art models, and introduce a new architecture, factorAtn, that shows improved robustness.
What Problem Does This Paper Address?
Deep learning has revolutionized image compression, achieving impressive rate-distortion performance (smaller file sizes for a given quality). However, as these models become more sophisticated, their susceptibility to adversarial attacks—subtle input modifications designed to fool the model—has become a growing concern. This paper specifically addresses the bitrate robustness of learned image compressors. The problem is that imperceptible changes to an input image can cause these models to produce significantly larger compressed representations, potentially leading to denial-of-service (DoS) attacks by exhausting storage or network bandwidth. This vulnerability has not been thoroughly investigated, leaving a critical gap in understanding the security and reliability of these advanced compression techniques.
Methods
The authors introduce two primary attack methodologies:
- White-box Attack: This attack assumes full knowledge of the target compression model's architecture and weights. It uses a gradient-based approach, specifically gradient ascent on the entropy estimation of the compressed latent representation. By maximizing this entropy, the attack effectively tricks the compressor into using more bits to encode the data, thus increasing the bitrate. The attack is constrained to ensure the input perturbations remain imperceptible.
- Black-box Attack: In this scenario, the attacker has no knowledge of the target model's internals. To overcome this, they propose a substitute model called DCT-Net. This network mimics JPEG compression by incorporating a Discrete Cosine Transform (DCT) layer, followed by quantization and a factorized entropy model. Adversarial images are generated using DCT-Nets trained with different quantization tables (mimicking different quality settings) and then transferred to the target learned image compressor. This method leverages adversarial transferability.
To improve robustness, the authors propose factorAtn, a novel compression architecture. It builds upon a basic factorized entropy model and incorporates attention modules into the analysis and synthesis transforms. This architecture aims to balance rate-distortion performance with enhanced resilience against adversarial attacks.
Datasets
The experiments were conducted using the Kodak dataset, a standard benchmark for image compression research, consisting of 24 images. The authors also mention using the Vimeo90K dataset for training their proposed factorAtn model and the substitute DCT-Nets.
Key Insights and Ideas
- Fragility of Learned Compression: State-of-the-art learned image compression models, despite their excellent rate-distortion performance, are surprisingly fragile when subjected to adversarial attacks designed to inflate bitrate.
- Entropy Estimation as an Attack Surface: The entropy estimation component of learned compressors is a critical vulnerability. Manipulating this estimation can lead to significant bitrate increases.
- Complex Models are Less Robust: Counter-intuitively, more complex entropy models (like hyperpriors) that achieve better rate-distortion performance tend to be less robust to these attacks compared to simpler models like the basic "factor" model.
- Attention as a Defense Mechanism: Incorporating attention modules into the analysis and synthesis transforms of image compression models appears to enhance their robustness against adversarial attacks.
- Adversarial Images and Distortion: While the primary goal of the attack is bitrate inflation, the paper notes that adversarial images typically introduce minimal reconstruction distortion, though outliers can occur.
Key Results
- White-box Attack: Achieved bitrate increases of up to 56.326× on the
hyperCmmodel at quality level 1. Across various models and qualities, significant bitrate inflation was observed. - Black-box Attack: Achieved bitrate increases of up to 1.947× on the
hyperCmmodel at quality level 6. The transferability of attacks from DCT-Nets proved effective. factorAtnPerformance: The proposedfactorAtnmodel demonstrated a promising trade-off between rate-distortion performance and robustness, outperforming existing models in both white-box and black-box adversarial settings, particularly at lower quality levels.- JPEG Robustness: Conventional JPEG compression showed greater robustness to adversarial attacks in the black-box setting compared to most learned image compression models, even those with similar bitrate qualities.
- Random Noise vs. Adversarial Perturbations: Learned image compressors were found to be relatively robust to random noise, with minimal impact on bitrate, unlike adversarial perturbations.
Limitations and Future Work
- Model Architecture Homogeneity: The study primarily focused on autoencoder-based architectures. Future work could explore attacks on other compression paradigms like RNNs or GANs.
- Network Capacity: The effect of varying network capacity (number of layers/channels) on robustness was not investigated.
- Distortion Metric: The current attack loss function prioritizes bitrate increase and does not strictly enforce minimal reconstruction distortion. Future work could incorporate additional loss terms for reconstruction quality.
- Higher Bitrate Qualities: The authors suggest extending attacks to models with even higher bitrate qualities.
- Alternative Distortion Metrics: Exploring attacks with different distortion metrics (e.g., MS-SSIM) is a potential avenue.
- Other Datasets: Evaluating attacks on different datasets beyond Kodak is also suggested.
Summary updated