RoVISQ: Reduction of Video Service Quality via Adversarial Attacks on Deep Learning-based Video Compression

Jung-Woo Chang, Mojan Javaheripi, Seira Hidano, Farinaz Koushanfar

Abstract

Video compression plays a crucial role in video streaming and classification systems by maximizing the end-user quality of experience (QoE) at a given bandwidth budget. In this paper, we conduct the first systematic study for adversarial attacks on deep learning-based video compression and downstream classification systems. Our attack framework, dubbed RoVISQ, manipulates the Rate-Distortion ($R-D$) relationship of a video compression model to achieve one or both of the following goals: (1) increasing the network bandwidth, (2) degrading the video quality for end-users. We further devise new objectives for targeted and untargeted attacks to a downstream video classification service. Finally, we design an input-invariant perturbation that universally disrupts video compression and classification systems in real time. Unlike previously proposed attacks on video classification, our adversarial perturbations are the first to withstand compression. We empirically show the resilience of RoVISQ attacks against various defenses, i.e., adversarial training, video denoising, and JPEG compression. Our extensive experimental results on various video datasets show RoVISQ attacks deteriorate peak signal-to-noise ratio by up to 5.6dB and the bit-rate by up to ~2.4× while achieving over 90% attack success rate on a downstream classifier. Our user study further demonstrates the effect of RoVISQ attacks on users' QoE.

Introduction

Video content accounts for over 80% of global internet traffic. To manage this massive load, the industry is transitioning toward Deep Learning (DL)-based video compression, which significantly outperforms traditional codecs like H.264. However, the security of these neural compression pipelines remains largely unexplored. While adversarial attacks on image classifiers are well-documented, video presents a unique challenge: lossy compression. Standard adversarial perturbations are typically "washed out" by the compression process, which treats them as high-frequency noise to be discarded.

Comparison of a benign video compression pipeline versus the RoVISQ attack framework, which degrades visual quality, bloats bandwidth, and sabotages downstream recognition.

RoVISQ (Reduction of Video Service Quality) is the first systematic framework designed to craft perturbations that not only survive compression but exploit the encoder's logic to degrade service quality or cause misclassification in downstream tasks like activity recognition.

Methodology: Manipulating the R-D Curve

RoVISQ treats the video compression pipeline as a differentiable optimization problem. Traditional encoders aim to minimize distortion ($D$) for a given bit-rate ($R$). RoVISQ subverts this fundamental trade-off through three primary attack modes:

  1. Video Quality Attack: Maximizes $D$ while maintaining a constant $R$, resulting in a severely degraded visual experience.
  2. Bandwidth Attack: Maximizes $R$ while keeping $D$ constant. This forces the encoder to spend extra bits on invisible noise, potentially exceeding bandwidth budgets and causing a Denial-of-Service (DoS) through buffering.
  3. RD Attack: Simultaneously maximizes both distortion and bit-rate for maximum disruption.
The effect of RoVISQ on the Rate-Distortion (R-D) curve, showing how attacks shift the operating point from the optimal curve (A) toward higher distortion (B), higher bit-rate (C), or both (D).

To ensure effectiveness across video sequences, RoVISQ incorporates Temporal Awareness, modeling how errors propagate from reference I-frames to predicted P/B-frames within a Group of Pictures (GOP). For real-time applications, the framework utilizes Universal Perturbations—pre-trained, input-invariant noise patterns that can be injected into any live stream without prior knowledge of the content.

Experimental Evaluation and Key Results

The researchers evaluated RoVISQ using industry-standard datasets, including Vimeo-90K for training universal perturbations and HEVC Test Sequences for performance metrics. Downstream classification was tested on UCF-101 (action recognition) and 20BN-JESTER (gesture recognition).

  • Visual and Network Impact: The attacks reduced the Peak Signal-to-Noise Ratio (PSNR) by up to 5.6dB and increased the bit-rate by up to 2.4×. In practice, this turns a high-definition stream into a blocky, buffering experience.
  • Classifier Sabotage: RoVISQ achieved over a 90% success rate in fooling downstream AI classifiers. Crucially, these attacks remained robust even when subjected to defenses like JPEG compression or video denoising.
  • User Experience: A study with 55 participants confirmed that RoVISQ significantly degrades the Quality of Experience (QoE), with many users indicating they would abandon a service under such attack conditions.

Conclusion

RoVISQ exposes a critical vulnerability in the next generation of video infrastructure. By making adversarial noise appear as "important" structural information, it forces neural encoders to prioritize the preservation of malicious noise over actual content. While currently demonstrated as a digital-domain attack, future work will explore physical-world triggers and the development of compression-aware adversarial training to secure DL-based codecs.

Summary updated

Abstract

Video compression plays a crucial role in video streaming and classification systems by maximizing the end-user quality of experience (QoE) at a given bandwidth budget. In this paper, we conduct the first systematic study for adversarial attacks on deep learning-based video compression and downstream classification systems. Our attack framework, dubbed RoVISQ, manipulates the Rate-Distortion ($\textit{R}$-$\textit{D}$) relationship of a video compression model to achieve one or both of the following goals: (1) increasing the network bandwidth, (2) degrading the video quality for end-users. We further devise new objectives for targeted and untargeted attacks to a downstream video classification service. Finally, we design an input-invariant perturbation that universally disrupts video compression and classification systems in real time. Unlike previously proposed attacks on video classification, our adversarial perturbations are the first to withstand compression. We empirically show the resilience of RoVISQ attacks against various defenses, i.e., adversarial training, video denoising, and JPEG compression. Our extensive experimental results on various video datasets show RoVISQ attacks deteriorate peak signal-to-noise ratio by up to 5.6dB and the bit-rate by up to $\sim$ 2.4$\times$ while achieving over 90$\%$ attack success rate on a downstream classifier. Our user study further demonstrates the effect of RoVISQ attacks on users' QoE.
BibTeX
@inproceedings{chang2023rovisq,
  author = {Jung-Woo Chang and Mojan Javaheripi and Seira Hidano and Farinaz Koushanfar},
  title = {{RoVISQ: Reduction of Video Service Quality via Adversarial Attacks on Deep Learning-based Video Compression}},
  booktitle = {NDSS},
  year = {2023},
  url = {https://openreview.net/forum?id=fNW-IjIJ1E},
}
Figure