When Authentication Is Not Enough: Breaking Behavior-Based Driver Authentication Systems

This paper addresses critical security and practical implementation gaps in existing behavioral-based driver authentication systems, which are increasingly driven by Artificial Intelligence (AI) for enhanced vehicle security. While these models propose powerful methods to identify drivers through unique biometric behavior, they have been insufficiently scrutinized from a security perspective regarding their connection to vehicle networks and vulnerability to adversarial attacks. The authors bridge this gap by proposing the first security-aware system model for behavioral-based driver authentication, developing lightweight architectures and novel evasion attacks, thereby aiding practitioners in safely adopting these systems.

System Model and Architectures

The research proposes two lightweight driver authentication systems designed for constrained environments: a Random Forest (RF) architecture and a single-layer Gated Recurrent Unit (GRU) architecture. These models are engineered to be efficient while outclassing the state-of-the-art, achieving an accuracy of up to 0.999 on real driving data. The system model formalizes a realistic vehicle network threat model, focusing on the security implications of deploying these systems within a real-world automotive context.

The key components of the proposed system model include:

((

(1) We propose the first security-aware system model for behavioral-based driver identification and authentication. Based on our design, we develop two new lightweight behavior-based driver authentication and identification systems whose requirements are compatible with commercial vehicle networks. Our two proposed architectures, i.e., a Random Forest (RF) and a single-layer Gated Recurrent Unit (GRU), are designed to be efficient in a realistic system model and outclass the state-of-the-art, achieving an accuracy of up to 0.999.

Data Collection and Processing

The data collection technique involves retrieving messages directly from the CAN bus, assuming the authenticator has already been equipped with the necessary information to decode messages and know which IDs are associated with which ECUs. Data is collected periodically:

1. For DL models leveraging causality between samples, data is aggregated in time windows of 16 seconds with a step size of 8. Time windows are then aggregated in batches with a size of 4, resulting in a prediction every 40 seconds.

2. For classical ML architecture, a prediction is generated for each collected sample (i.e., each second).

The dataset utilized is the widely adopted OCSLab dataset, which comprises 94,380 data points from 10 different drivers and 54 distinct features extracted from CAN bus messages. Preprocessing steps include normalization and undersampling to rebalance class distribution.

Threat Model and Evasion Attacks

The paper introduces a realistic threat model where potential attackers can physically access the vehicle’s CAN bus network, deploying a covert malicious device to inject packets. The attacker is assumed to know the DBC file of the vehicle and the features used by the authentication model. To mitigate safety risks, only 22 out of 46 features are identified as modifiable without affecting vehicle behavior (e.g., engine coolant temperature, intake air pressure).

Two novel evasion attacks are proposed:

1. SMARTCAN: This attack assumes the attacker has access to legitimate user data but not the authenticator model implementation. It uses a smart-replay attack by replaying legitimate traffic using only modifiable features while stealing the car, allowing authentication despite interference in driving safety.

2. GANCAN: This attack assumes the attacker has access to the authenticator model response but not legitimate driving behavior. It employs Reinforcement Learning (RL) to optimize a generator that crafts legitimate fake packets starting from noise, overwriting only modifiable features while non-modifiable and borderline features are extracted from the attacker’s own driving behavior.

Evaluation and Key Findings

The systems were evaluated using True Positive (TP), False Positive (FP), False Negative (FN), and True Negative (TN) metrics, along with Accuracy, F1 score, and Attack Success Rate (ASR).

(2) We introduce SMARTCAN and GANCAN, the first attacks against behavioral-based driver authentication and identification systems. Our attacks use evasion techniques to avoid detection and only inject minimal amounts of data to preserve driving functions while achieving success rates of up to 1.000.

Baseline evaluations showed that our models can outclass the state-of-the-art, with the RF model achieving an accuracy of up to 0.998 in identification and authentication. The paper introduces combinatorial accuracy, a concept that reduces false positive alerts by waiting for multiple consecutive decisions before triggering a notification, which significantly lowers the probability of false alarms at the cost of a couple of seconds of delay.

Takeaways for Practitioners

The paper concludes with five key takeaways aimed at bridging the gap between research and practice:

1. Behavioral-based driver authentication systems should be implemented as ECUs in the CAN bus to reduce the probability of tampering from malicious parties.

Based on the provided research paper, here are specific improvements that can be made to existing behavioral-based driver authentication systems, and what those improved AI systems could achieve:

I. System Architecture and Deployment Improvements (Addressing RQ1 & RQ4)

1. A paradigm shift in deployment from external OBD-II port devices to an integrated ECU implementation directly on the CAN bus.

2. Integration of a robust CAN message authentication protocol (e.g., leveraging AUTOSAR SecOC) as a fundamental layer of security for all vehicle network communications, rather than relying solely on behavioral pattern recognition for security.

3. The AI system will transition from being just an identification tool to acting as a continuous, real-time anomaly detection and intrusion prevention system embedded within the vehicle's control logic.

4. By implementing this architecture, the AI system can achieve:

5. Continuous, tamper-resistant monitoring of driver behavior at the lowest possible network level without requiring external hardware access or relying on vulnerable external ports like OBD-II.

II. Model Architecture and Efficiency Improvements (Addressing RQ2)

1. The adoption of lightweight, resource-efficient architectures, specifically favoring single-layer Gated Recurrent Unit (GRU) networks over deeper or more complex models, while retaining high accuracy (up to 0.999).

2. Implementation of specialized feature selection techniques using Explainable AI (XAI) methods like SHAP values to dynamically identify and prioritize the most critical CAN features for authentication, reducing computational load without sacrificing security.

3. The improved system will be able to:

4. Provide high-accuracy driver identification and authentication in real-time on resource-constrained in-vehicle ECUs (like a Raspberry Pi 4B), ensuring continuous operation without constant reliance on cloud resources or powerful external GPUs.

5. Ensure user privacy by mandating local training and inference, meaning sensitive driving behavior data never leaves the vehicle's secure environment.

III. Robustness Against Adversarial Attacks (Addressing RQ4)

1. The system must be hardened against evasion attacks such as SMARTCAN (data replay) and GANCAN (model poisoning/generation).

2. The improved AI will incorporate adversarial training techniques, optimizing the generator models to bypass fixed discriminators rather than relying on simple incremental training, thereby increasing resilience against sophisticated data manipulation.

3. This hardened system will be able to:

4. Maintain a near-perfect security posture (ASR close to 1.0) even when an attacker has knowledge of the target model or data (GANCAN scenario), ensuring that malicious packet injection is infeasible or extremely time-consuming (requiring significant offline training time).

5. The system will be able to:

6. Detect and mitigate attempts to impersonate legitimate drivers by analyzing feature interactions and temporal dependencies, distinguishing between genuine behavioral evolution and malicious data injections.

IV. Operational Reliability and Usability Improvements (Addressing RQ3)

1. The implementation of a Combinatorial Accuracy mechanism, using the geometric distribution formula to require multiple consecutive unauthorized batch detections before triggering an alert.

2. The improved system will be able to:

3. Drastically reduce false positive rates (dropping by orders of magnitude) while maintaining high security, providing a fair trade-off between safety alerts and user experience.

4. The final system will be capable of:

5. Providing actionable, low-false-alarm notifications to the vehicle owner only after multiple independent behavioral anomalies are confirmed, ensuring that legitimate drivers have an uninterrupted driving experience while still protecting the vehicle from theft or misuse.

V. Strategic Positioning Improvements (Addressing RQ5)

1. The system will be strategically positioned as a second-factor continuous authentication mechanism layered on top of traditional physical keys, rather than replacing them as the sole means of access.

2. The improved AI system will be capable of:

3. Detecting and reporting unauthorized use or theft attempts even after an initial breach (e.g., key compromise), acting as a persistent security layer that monitors driving state for potential fraud or unauthorized operation, thereby enhancing overall vehicle security posture without sacrificing the convenience of physical access.

Researchers extensively explored behavior-based driver authentication systems in vehicles. Pushed by advances in Artificial Intelligence (AI), these systems employ powerful models to identify drivers based on unique biometric behaviors. However, existing work prioritizes AI performance metrics, neglecting secure integration with real-world automotive environments and the threat of adversarial attacks that can fool the authentication system. In this paper, we propose for the first evasion attacks against behavior-based driver authentication systems, allowing an attacker to impersonate the legitimate driver. Our attacks exploit long-standing CAN bus weaknesses that allow the injection of forged frames without jeopardizing the attacker's safety while stealing the vehicle. When legitimate data samples are available, we propose SMARTCAN, a safety-aware replay attack. If the attacker can only use the authenticator as an oracle, we propose GANCAN, which trains a Generative Adversarial Network's generator using reinforcement learning on the authenticator's responses. Our attacks achieve a success rate up to 100% against all the considered models and, in the worst case, require 22 minutes to steal a vehicle. Acknowledging our identified vulnerabilities, we discuss the requirements for a safe and effective deployment of these systems in real-world scenarios.