Edge-Semantic Perceiver IO for Firmware Backdoor Detection Using Binary Function Call Graphs

Abstract

Firmware backdoors are difficult to detect after compilation because malicious behavior may be confined to a small set of functions and remain hidden after extensive changes to instruction layout. This paper proposes ESPI-FBD, an edge-semantic Perceiver IO framework that analyzes binary function call graphs while avoiding quadratic interaction among all recovered functions. The method integrates architecture-normalized function descriptors, typed directional call-relation biases, salience-seeded latent vectors, latent self-attention, and separate output queries for firmware classification and suspicious-function ranking. A product-disjoint evaluation protocol is specified for a multi-architecture firmware corpus and covers detection, localization, cross-architecture transfer, robustness, calibration, latency, and memory. The framework is designed to preserve sparse malicious evidence in a bounded latent workspace and to return an analyst-oriented review queue in addition to an image-level decision. The numerical values currently reported are illustrative placeholders that define the intended evaluation and must be replaced with audited experimental outputs before submission.