Robust algorithms for image compression and decompression and their trade-offs

Research Themes

Sensor Signal Processing

Aim

Our aim is to study trade-offs in image compression algorithms between robustness and several task-specific metrics, like rate, classification error, perception quality, and reconstruction performance, in order to gain insights to develop robust algorithms, with application in extreme compression scenarios, like underwater communication.

Objectives

1. Develop a theoretical characterisation of the tradeoffs in image compression between robustness, compression rate, classification error, perception quality, and reconstruction performance.
2. Design new image compression algorithms that select features according to the task at end and that are robust to adversarial attacks. 
3. Validate the algorithms in realistic scenarios, where images undergo extreme compressed due to limitations in bandwidth, as in underwater communication.

Description

Deep neural networks (DNNs) have become essential tools in automated decision-making, powering applications from image classification and segmentation to anomaly detection and portfolio allocation. However, DNNs are notoriously vulnerable to adversarial attacks, prompting extensive research over the past decade to develop new attacks and defense strategies [1-4]. Whilst implementing these defenses can reduce performance on non-perturbed samples [5-7], their impact on multi-task performance remains unclear.

In this project, we explore image compression and decompression algorithms where the target image serves various tasks, such as classification, segmentation, or multimedia. By leveraging knowledge of these downstream tasks, the encoder can select better features and achieve improved compression rates [8]. However, there are inherent trade-offs between compression rates and task performance [9]. Our goal is to study how adversarial attacks influence these tradeoffs, design robust mechanisms for the entire pipeline, for example, leveraging semantically-meaningful representations, and demonstrate the application of these algorithms in scenarios requiring extreme compression ratios, such as underwater communication.

[1] C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, R. Fergus, "Intriguing Properties Of Neural Networks," arXiv:1312.6199v4, 2014.

[2] N. Akhtar, A. Mian, N. Kardan, and M. Shah, ‘Advances in Adversarial Attacks and Defenses in Computer Vision: A Survey’, IEEE Access, vol. 9, pp. 155161–155196, 2021, doi: 10.1109/ACCESS.2021.3127960.

[3] A. Modas, R. Sanchez-Matilla, P. Frossard, and A. Cavallaro, ‘Toward Robust Sensing for Autonomous Vehicles: An Adversarial Perspective’, IEEE Signal Process. Mag., vol. 37, no. 4, pp. 14-23, Jul. 2020, doi: 10.1109/MSP.2020.2985363.

[4] D. Hendrycks, S. Basart, N. Mu, S. Kadavath, F. Wang, E. Dorundo, R. Desai, T. Zhu, et al., "The Many Faces Of Robustness: A Critical Analysis Of Out-of-Distribution Generalization," ICCV, pp. 8340-8349, 2021.

[5] D. Tsipras, S. Santurkar, L Engstrom, A. Turner, A. Madry, "Robustness May Be at Odds with Accuracy," ICLR, 2019.

[6] H. Zhang, Y. Yu, J. Jiao, E. Xing, L. El Ghaoui, M. Jordan, "Theoretically Principled Tradeoff Between Robustness and Accuracy," ICML, PMLR 97:7472-7482, 2019.

[7] M. Mehrabi, A. Javanmard, R. A. Rossi, A. Rao, T. Mai, "Fundamental Tradeoffs in Distributionally Adversarial Training," ICML, 2021.

[8] Z. Lei, P. Duan, X. Hong, J. F. C. Mota, J. Shi, and C.-X. Wang, "Progressive Deep Image Compression for Hybrid Contexts of Image Classification and Reconstruction", IEEE J. Select. Areas Commun., vol. 41, no. 1, pp. 72–89, Jan. 2023, doi: 10.1109/JSAC.2022.3221998.

[9] J. Fang, J. F. C. Mota, B. Lu, W. Zhang, and X. Hong, "The Rate-Distortion-Perception-Classification Tradeoff: Joint Source Coding and Modulation via Inverse-Domain GANs", IEEE Trans. Signal Process., vol. 72, pp. 3076-3090, 2024, doi: 10.1109/TSP.2024.3411692.

[10] A. Wheeldon and A. Serb, "A study on the clusterability of latent representations in image pipelines," Front. Neuroinform, vol. 17, no. 1074653, pp. 1-11, 2023. doi: 10.3389/fninf.2023.1074653

Minimum entry qualification - an Honours degree at 2:1 or above (or International equivalent) in a relevant science or engineering discipline, possibly supported by an MSc Degree. Further information on English language requirements for EU/Overseas applicants.

Home rate fees and stipend are available for this position.