TL;DR
This paper introduces WattsUpDoc, a novel system that uses power consumption side channels to detect malware on embedded medical devices without any hardware or software modifications. It achieves high accuracy, comparable to traditional antivirus, by analyzing subtle variations in power draw, offering a crucial security layer for devices that are otherwise difficult to protect.
What Problem Does This Paper Address?
Modern healthcare relies heavily on embedded medical devices, many of which are increasingly connected to networks. This connectivity, while beneficial, exposes these devices to malware. However, traditional security solutions like antivirus software and OS updates are often incompatible with these devices due to custom firmware, resource constraints, or manufacturers' restrictions that prevent modifications to maintain device certification. This creates a significant security gap, leaving critical medical equipment vulnerable and hindering the reporting of security incidents.
Methods
WattsUpDoc employs a nonintrusive approach by monitoring the device's power consumption at the wall outlet. This "side channel" leaks information about the device's internal activity. The core of WattsUpDoc is a machine learning system that uses supervised learning to distinguish between normal and abnormal power consumption patterns.
Key methodological aspects include:
- Power Monitoring: A simple, low-cost setup using a sense resistor to measure voltage drop across the neutral terminal of the power outlet, capturing AC power traces at a high sampling rate (250 kHz).
- Feature Extraction: Time-domain features (mean, variance, RMS, etc.) and frequency-domain features are extracted from 5-second windows of the power traces. These features capture variations in workload and system activity.
- Supervised Learning: Three classifiers (3-Nearest Neighbors, Multilayer Perceptron, and Random Forest) are trained on labeled data of normal and abnormal (malware-infected) device behavior.
- Behavioral Analysis: Instead of relying on malware signatures, WattsUpDoc learns the "normal" operational profile of a device and flags deviations as potentially malicious. This is particularly effective for embedded systems with constrained, repetitive tasks.
Datasets
The paper evaluates WattsUpDoc on two types of embedded devices:
- Baxa ExactaMix 2400 Pharmaceutical Compounder: Runs Windows XP Embedded.
- Schweitzer SEL3354 Substation Computer: A SCADA device running Windows XP Embedded.
The authors collected AC power traces for these devices under various conditions, including normal operation (idle, booting, specific tasks) and infected with both real-world and emulated malware samples. The malware samples used are listed in Table 4 and represent common untargeted threats. The datasets were partitioned for training and testing, with experiments conducted using both known and unknown malware samples.
Key Insights and Ideas
- Power Consumption as a Rich Security Signal: The paper demonstrates that seemingly innocuous power fluctuations can be highly indicative of complex software activities, including malware execution. This is especially true for embedded systems with predictable operational states.
- Nonintrusiveness is Key: The ability to detect malware without any hardware or software modifications is a significant advantage for medical devices, where such changes are often prohibited or prohibitively expensive.
- Behavioral Profiling Outperforms Signatures for Embedded Systems: For devices with limited state spaces and repetitive functions, learning a baseline of "normal" behavior is more robust than relying on signature-based detection, which can be easily evaded or requires constant updates.
- Untargeted Malware is a Real Threat: The study focuses on common malware, highlighting that even without sophisticated, targeted attacks, a significant risk exists for connected embedded devices.
- Feature Importance Varies by Device: For the compounder, features related to the dynamic range of power draw (due to pumps) were most important. For the substation computer, frequency-domain features indicating rapid power changes were more significant, reflecting its different operational characteristics.
Key Results
- High Accuracy on Known Malware: WattsUpDoc achieved accuracies of 94% on the compounder and 99% on the substation computer for detecting malware it was trained on.
- Strong Performance on Unknown Malware: For malware samples not included in the training set, accuracy was 88.5% on the compounder and 84.9% on the substation computer, which is comparable to state-of-the-art behavior-based malware detectors.
- Low False Positives: Precision rates were consistently high (e.g., 93.5% on the compounder for unknown malware), indicating that flagged events are highly likely to be actual infections, minimizing unnecessary downtime.
- Soft Real-Time Capability: The system can classify events with a 5-second lag, processing 5-second chunks of data and performing feature extraction and classification within that window.
- Optimal Window Size: A 5-second window for feature calculation provided the highest accuracy across both devices.
- Training Set Size: Diminishing returns in accuracy were observed after approximately 500 training samples for the compounder and 1000 for the substation computer.
Limitations and Future Work
- Generalizability to PCs: The authors note that WattsUpDoc's effectiveness is tied to the constrained state space of embedded systems. It may struggle with general-purpose PCs due to their vast number of possible states and concurrent processes, potentially leading to false alarms.
- Untargeted Malware Focus: The system is designed to detect untargeted malware and does not address sophisticated, targeted attacks that might specifically evade power analysis.
- Deployment Complexity: While the measurement points are simple, managing and analyzing data from numerous devices in a large network could still pose a challenge.
- Balancing False Positives/Negatives: The paper acknowledges that finding the optimal balance between detecting all malware (high recall) and minimizing false alarms (high precision) is an ongoing challenge for practical deployment.
- Future Work: The authors suggest exploring complementary NIDS integration, developing more compact measurement hardware (e.g., outlet covers, power bricks), and further investigating the generalizability to devices with firmware rather than full operating systems.
Summary updated