NIC-RobustBench: A Comprehensive Open-Source Toolkit for Neural Image Compression and Robustness Analysis

Georgii Bychkov, Khaled Abud, Egor Kovalev, Aleksandr Gushchin, Sergey Lavrushkin, Dmitriy S. Vatolin, Anastasia Antsiferova

Keywords: image сompression, adversarial robustness, open-source model, algorithm library

TL;DR: This paper introduces the first open-source framework for evaluating the adversarial robustness of neural image compression methods

Abstract: Adversarial robustness of neural networks is a rapidly developing and important research area, ensuring the reliable use of neural network models in areas such as computer vision, natural language processing, and others. With the emergence of neural image compression (NIC) methods and the introduction of the first standard, JPEG AI, the question of robustness has also become critical in image compression. Unstable NIC models are more prone to producing distortions and can be exploited to compromise downstream vision tasks that rely on compressed data. Most existing benchmarks for NIC focus primarily on compression efficiency and reconstruction quality, while prior robustness studies have been limited to a small number of codecs and attacks. To address this gap, we introduce \textbf{NIC-RobustBench}, the first open-source framework for evaluating the adversarial robustness of NIC methods. It integrates 6 adversarial attacks and 7 defense strategies in addition to traditional Rate-Distortion (RD) evaluation. The framework includes the largest number of codec types among all known NIC libraries and is easily scalable. The paper provides a comprehensive overview of NIC-RobustBench alongside an extensive robustness evaluation of modern NIC methods. Our code is publicly available at \textit{link hidden for blind review}. We believe our framework will become an essential tool for developing robust neural image compression techniques.

Primary Area: infrastructure, software libraries, hardware, systems, etc.

Submission Number: 24709