arXiv subjects

Aurélien Francillon

Publications and source records attributed to Aurélien Francillon.

4 recordsLinked to original sources

arXiv2026

Verifiable and Confidential DNN Inference on Low-End Edge Devices

Deploying deep neural network (DNN) inference on low-end edge devices raises two key challenges: protecting model confidentiality against a potentially compromised edge system and enabling verifiable inference without incurring prohibitive overhead. Existing approaches either house partial models and inference software within trusted execution environments (TEEs), resulting in high cost and an application-dependent trusted computing base (TCB), or execute in untrusted environments, providing little security. In this work, we present VECODI, a framework for verifiable and confidential DNN inference on constrained edge devices. At its core, VECODI introduces SHANGRI-LA, a new execution abstraction on TrustZone-M TEEs that establishes a third runtime environment with privileges strictly between the Secure and Non-Secure Worlds. VECODI leverages SHANGRI-LA to execute untrusted inference code in the Non-Secure World while using minimal application-agnostic Secure-World support to protect model confidentiality and enable verifiability (with respect to proper execution of inference code and model parameters) of inference results. We realize VECODI on a real-world NUCLEO-L552ZE-Q development board and open-source its prototype. Our results show VECODI's small TCB, memory footprint, and runtime overhead, making it a practical option for secure inference in low-end edge devices.

cs.CR↗

arXiv2026

Evasive Intelligence: Lessons from Malware Analysis for Evaluating AI Agents

Artificial intelligence (AI) systems are increasingly adopted as tool-using agents that can plan, observe their environment, and take actions over extended time periods. This evolution challenges current evaluation practices where the AI models are tested in restricted, fully observable settings. In this article, we argue that evaluations of AI agents are vulnerable to a well-known failure mode in computer security: malicious software that exhibits benign behavior when it detects that it is being analyzed. We point out how AI agents can infer the properties of their evaluation environment and adapt their behavior accordingly. This can lead to overly optimistic safety and robustness assessments. Drawing parallels with decades of research on malware sandbox evasion, we demonstrate that this is not a speculative concern, but rather a structural risk inherent to the evaluation of adaptive systems. Finally, we outline concrete principles for evaluating AI agents, which treat the system under test as potentially adversarial. These principles emphasize realism, variability of test conditions, and post-deployment reassessment.

cs.CR↗

arXiv2015

Automated Dynamic Firmware Analysis at Scale: A Case Study on Embedded Web Interfaces

Embedded devices are becoming more widespread, interconnected, and web-enabled than ever. However, recent studies showed that these devices are far from being secure. Moreover, many embedded systems rely on web interfaces for user interaction or administration. Unfortunately, web security is known to be difficult, and therefore the web interfaces of embedded systems represent a considerable attack surface. In this paper, we present the first fully automated framework that applies dynamic firmware analysis techniques to achieve, in a scalable manner, automated vulnerability discovery within embedded firmware images. We apply our framework to study the security of embedded web interfaces running in Commercial Off-The-Shelf (COTS) embedded devices, such as routers, DSL/cable modems, VoIP phones, IP/CCTV cameras. We introduce a methodology and implement a scalable framework for discovery of vulnerabilities in embedded web interfaces regardless of the vendor, device, or architecture. To achieve this goal, our framework performs full system emulation to achieve the execution of firmware images in a software-only environment, i.e., without involving any physical embedded devices. Then, we analyze the web interfaces within the firmware using both static and dynamic tools. We also present some interesting case-studies, and discuss the main challenges associated with the dynamic analysis of firmware images and their web interfaces and network services. The observations we make in this paper shed light on an important aspect of embedded devices which was not previously studied at a large scale. We validate our framework by testing it on 1925 firmware images from 54 different vendors. We discover important vulnerabilities in 185 firmware images, affecting nearly a quarter of vendors in our dataset. These experimental results demonstrate the effectiveness of our approach.

cs.CR↗

arXiv2009

Code injection attacks on harvard-architecture devices

Harvard architecture CPU design is common in the embedded world. Examples of Harvard-based architecture devices are the Mica family of wireless sensors. Mica motes have limited memory and can process only very small packets. Stack-based buffer overflow techniques that inject code into the stack and then execute it are therefore not applicable. It has been a common belief that code injection is impossible on Harvard architectures. This paper presents a remote code injection attack for Mica sensors. We show how to exploit program vulnerabilities to permanently inject any piece of code into the program memory of an Atmel AVR-based sensor. To our knowledge, this is the first result that presents a code injection technique for such devices. Previous work only succeeded in injecting data or performing transient attacks. Injecting permanent code is more powerful since the attacker can gain full control of the target sensor. We also show that this attack can be used to inject a worm that can propagate through the wireless sensor network and possibly create a sensor botnet. Our attack combines different techniques such as return oriented programming and fake stack injection. We present implementation details and suggest some counter-measures.

cs.CR↗