IoT Firmware Version Identification Using Transfer Learning with Twin Neural Networks

Ashley Andrews, George Oikonomou, Simon Armour, Paul Thomas, Thomas Cattermole · 2025

Read IoT Firmware Version Identification Using Transfer Learning with Twin Neural Networks by Ashley Andrews, George Oikonomou, Simon Armour, Paul Thomas, Thomas Cattermole (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-06033.txt.

Introduction

Ashley Andrews and co-authors spot firmware updates in smart devices from their network traffic alone, in a research paper from 2025. Keeping home devices on the latest firmware patches known holes, so knowing when a version changes is a security plus. Version shifts leave only faint traces in traffic, and training data are scarce.

Their trick uses a twin neural network first trained to tell devices apart. Traffic snapshots become gray images, and the twin scores how alike two snapshots look. A standard effect size on those scores flags a version change. In a lab with twelve devices watched for eleven days, the best model caught stable versions at 95.83 percent and changes at 84.38 percent.

Catalog record

CREATORAshley Andrews, George Oikonomou, Simon Armour, Paul Thomas, Thomas Cattermole
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE69 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-06033

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

IoT Firmware Version Identification Using Transfer Learning with Twin Neural Networks Ashley Andrews, George Oikonomou, Simon Armour, Paul Thomas, Thomas Cattermole Abstract . As the Internet of Things (IoT) becomes more embedded within our daily lives, there is growing concern about the risk ‘smart’ devices pose to network security. To address this, one avenue of research has focused on automated IoT device identification. This research is broadly motivated by the idea that the more we can know about our devices, the more secure the networks they are on can be. Research has however largely neglected the identification of IoT device firmware versions. There is strong evidence that IoT security relies on devices being on the latest version patched for known vulnerabilities. Identifying when a device has updated (has changed version) or not (is on a stable version) is therefore useful for IoT security. Version identification involves challenges beyond those for identifying the model, type, and manufacturer of IoT devices. Most obviously, the differences between versions are more subtle and therefore harder to detect. Moreover, because there has been relatively little research in this area, there are no widely available datasets that track devices’ version changes over time. Consequently, traditional machine learning algorithms are ill-suited for effective version identification due to being limited by the availability of data for training. In this paper, we introduce an effective technique for identifying IoT device versions based on transfer learning. This technique relies on the idea that we can use a Twin Neural Network (TNN) — trained at distinguishing devices — to detect differences between a device on different versions. This facilitates real-world implementation by requiring relatively little training data. In more detail, we extract statistical features from on-wire packet flows, convert these features into greyscale images, pass these images into a TNN to output similarity scores, and determine version changes based on the Hedges’ g effect size of the similarity scores. This allows us to detect the subtle changes present in on-wire traffic when a device changes version. To evaluate our technique, we set up a lab containing 12 IoT devices and recorded their on-wire packet captures for 11 days across multiple firmware versions. For testing data held out from training, our best performing model is shown to be 95.83% and 84.38% accurate at identifying stable versions and version changes respectively. DOI: XXXXXXX.XXXXXXX ISBN: 978-1-4503-XXXX-X/18/06 CCS: Security and privacy Network security CCS: Computing methodologies Neural networks 1. Introduction The impact of the Internet of Things (IoT) is increasingly being felt across all aspects of society: in our homes (so-called ‘smart homes’), businesses (industrial IoT), and the public sector (smart cities). As more IoT devices become embedded within our daily lives, there is a growing fear about the potential risk they pose to security (not to mention privacy and safety). These fears have been realised with “smart doorbells being an ‘easy target for hackers’" ( BBC, 2020 ) and seemingly innocuous children’s toys allowing attackers access into networks and subsequently homes ( BBC, 2017 ) . To combat this risk, a body of research is developing devoted to automatically identifying IoT devices and understanding their behaviour. Recent work has largely focused on using supervised Machine Learning (ML) to classify wired (and wireless) network traffic in terms of device models, types and manufacturers. Underlying this work is the broad motivation that the more we can know about our IoT devices automatically, the more we can secure the IoT. More specific motivations include: automatically detecting when a rogue device enters a network and automatically detecting when a known device is behaving anomalously and therefore might be compromised. To fully identify an IoT device, there are four properties as defined by Meidan et al ( Meidan et al., 2020 ) . These are: 1) Type, 2) Manufacturer, 3) Model Number and 4) Firmware Version. IoT device identification down to a software/firmware version level, specifically ensuring a device is running their latest version patched for known vulnerabilities ( Andrews et al., 2023c ) , is a motivation that has been largely neglected ( Andrews et al., 2023a ; Andrews et al., 2023b ) . In order to be competitive, device manufacturers need to continuously release new products with short time-to-market and lowest possible prices. To this end, in this specific market, it is extremely hard to produce devices with strong security guarantees ( Catuogno and Galdi, 2023 ) and therefore flaws and vulnerabilities in devices are prevalent. A recent case with a flaw in Hikvision cameras allowed remote access ( BBC, 2023 ) , showing the importance of applying firmware updates to keep devices secure. However even with a serious vulnerability and a patch available, a whitepaper published 12 months later by Cyfirma ( Cyfirma, 2022 ) showed the patches had not been applied to over 80,000 internet reachable devices. This neglected motivation therefore has a concrete real-world application for IoT security: if we can automate version identification we can keep IoT devices up to date without user intervention, reducing their burden. Version identification is however a more difficult problem to solve than other types of identification, such as device type or model. This is because the on-wire changes between versions are expected to be subtle and datasets are not publicly available containing version change information to train ML models. A method needs to be developed that can overcome these two challenges. This paper’s contributions include: • A technique using transfer learning to identify device version changes without a need to train on a large amount of data, or having seen the data previously. More specifically using the output similarity scores from Twin Neural Networks with Hedges’ g to identify both stable device versions and subtle device version changes, • Obtaining a best accuracy of 95.83% identifying stable versions and 84.38% identifying version changes on a real world test-bed when using Hedges’ g to identify subtleties on the Twin Neural Network output. The use of Hedges’ g gives an improvement of approximately 20% over alternative metrics, • A analysis of the difficulties in identifying IoT device firmware version changes seen in our testbed, specifically the subtle differences in on-wire signatures or in some cases no changes at all. Our structure for the paper is as follows. In Section 2 we give a clear real-world motivation for how automated version identification can improve IoT security. We also describe the challenges of version identification and outline a method for addressing these challenges. In Section 3 we examine the literature around IoT device identification and twin neural networks, including how we can interpret the output similarity scores when applied to IoT device versions. in Section 4 we then use this knowledge for our proposed method, extracting our chosen features to generate pairs to train and then test the TNN on two scenarios - identification of devices with a stable version and identification of device version changes - using four measures of accuracy. In Section 5 we briefly state the results from training the TNN and for our two test cases. We then complete a more in-depth analysis of the results with limitations of our approach, including analysis of which devices do and do not perform well. We state how these results

Read the complete text