Introduction
In the landscape of modern cybersecurity, traditional defenses often fall short against sophisticated, novel attacks. Presented at AAAI 2006, AWDRAT introduces a paradigm shift toward "cognitive immunity." Rather than merely blocking known threats, AWDRAT enables applications to sense incorrect behavior, diagnose the root cause of compromises, and adapt execution in real-time to maintain essential functionality.
The Survivability Problem
Current security relies on signature detection (identifying known "bad" patterns) and statistical anomaly detection (flagging "unusual" behavior). Both have significant flaws: signature-based systems are blind to zero-day exploits, while statistical systems often suffer from high false-positive rates, mistaking rare legitimate actions for attacks. Crucially, neither approach provides a path for system recovery post-breach. AWDRAT addresses this by moving from "detecting the bad" to "enforcing the good" through a model-based, self-healing architecture.
The AWDRAT Architecture
AWDRAT stands for Architectural-differencing, Wrappers, Diagnosis, Recovery, Adaptive software, and Trust-modeling. This middleware framework provides applications with self-awareness by comparing real-time execution against a formal model of intended behavior.
The system integrates several sophisticated components as shown in the architectural overview:
- System Architectural Models: Developers define how software should function, specifying pre-conditions, post-conditions, and invariants.
- Wrappers: Automatically generated sensors at the DLL and Java method levels feed execution data into the middleware.
- Trust Modeling: A persistent "trust score" is maintained for various components; if a resource is suspected of compromise, its score is downgraded.
- Adaptive Method Selection: Using decision theory, the system can "fail over" to alternative methods—for instance, switching from an untrusted native library to a secure Java implementation.
Architectural Differencing and Diagnosis
The core "cognitive" step in AWDRAT is architectural differencing. The system runs a simulation of the architectural model in parallel with the live application.
As illustrated in the differencing process, if the real output (e.g., a file write or network call) deviates from the simulated model's expectation, a conflict is flagged. AWDRAT then employs Model-Based Diagnosis—a technique borrowed from hardware troubleshooting—to identify which computational resources (code, data, or memory) are likely compromised. This diagnosis serves as the vital bridge between detection and intelligent recovery.
Evaluation and Results
The system was evaluated using the MAF (Mission-Builder), a DARPA interactive mission planner. Researchers subjected the system to 59 attack attempts, including manual code injections, parasitic threads, and malicious payloads delivered via corrupted image files.
The results demonstrated the robustness of the approach:
- 91% Detection Rate: AWDRAT successfully identified 54 out of 59 attacks.
- Zero False Positives: Because checks are performed against a deterministic model, legitimate user behavior was never incorrectly flagged.
- 100% Recovery in Specific Scenarios: In attacks involving malicious image files, the system achieved a perfect success rate in detecting, diagnosing, and correcting behavior by switching to safer loading methods.
- Negligible Overhead: For interactive applications, the performance impact of wrappers and monitoring did not degrade the user experience.
Conclusion and Future Work
While AWDRAT offers a powerful path toward survivable systems, it faces challenges regarding the initial modeling effort required for legacy software. Future work aims to improve thread awareness in the execution monitor and extend the framework to high-traffic server-side environments, such as Domain Name Servers (DNS). By enforcing intended behavior rather than just blocking known threats, AWDRAT provides a blueprint for resilient, self-healing software.
Summary updated