Architecture-Aware Neural Compression for TriCore Firmware Using Knowledge Distillation

View at publisher Summarise with VIRAT
Authors: Hyunjoong Lee, Hoseong Kim, Daejin Park
Cited by: 0
References: 9

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

The transition toward Software-Defined Vehicles (SDVs) has significantly increased the complexity and size of automotive ECU firmware, posing growing challenges in hardware cost (BOM) and over-the-air (OTA) update efficiency. Conventional general-purpose compressors fail to exploit the architectural characteristics of the TriCore processor, resulting in limited compression efficiency. To address this limitation, this paper proposes an AI-driven, architecture-aware compression framework tailored for TriCore-based firmware. The proposed method employs a two-stage knowledge distillation scheme, in which a large Transformer-based Teacher model transfers its learned representations of TriCore binary patterns to a lightweight GRU-based Student model suitable for on-device deployment in real ECUs. Experimental results show that the Student-based compression framework achieves an average compression efficiency of 1.524 bpb, corresponding to 19.05% of the original firmware size on the TC375 dataset, which represents an improvement of approximately 38% compared to conventional general-purpose compressors. In addition, the proposed approach exhibits a compact runtime memory footprint during decompression, requiring only 0.25 MB, which is approximately 94% lower than that of traditional methods. Furthermore, crossvalidation between the Python reference implementation and the C-based decoder confirms deterministic and consistent decoding behavior, demonstrating the practicality and portability of the proposed framework. Overall, this study establishes a strong proof of concept for AI-driven, architecture-aware firmware compression and highlights its potential for further optimization and deployment on embedded hardware.

Fields of Study & Subjects

Advanced Data Compression Techniques Algorithms and Data Compression Software Engineering Research

Publication details

Journal / source
International Conference on Artificial Intelligence in Information and Communication (ICAIIC)
Pages
937–941
Publication date
24 February 2026
Language
EN