Abstract
OACAL (Obligation, Attack scenario, model Checking And machine Learning) is a novel framework designed to automate the redesign of software specifications to ensure security even when users fail to adhere to security protocols, such as leaving a terminal unattended. By integrating module-based recomposition with machine learning-accelerated model checking, OACAL identifies functional and secure system designs up to 20 times faster than existing state-of-the-art methods while maintaining superior solution coverage.
The Problem: Weakened User Obligations
In security-critical systems, developers typically design workflows under the assumption that users will follow specific "obligations"—sequences of actions required to maintain safety. However, human behavior is often unpredictable; users may become distracted or deviate from intended workflows, creating vulnerabilities that malicious actors or automated threats like Remote Access Trojans (RATs) can exploit.
The challenge for developers is to make systems resilient to these "weakened obligations" by redesigning the system's specification—the logic governing state transitions—so it remains secure without sacrificing core functionality. Previous tools, such as OASIS, often struggle with slow processing speeds, limited search ranges, and the accidental removal of essential features during the redesign process.
The OACAL Framework
OACAL addresses these challenges through three primary innovations designed to streamline the search for secure specifications:
- Module-based Recomposition: Rather than shuffling individual low-level actions, OACAL groups actions into "modules" representing specific UI screens or interfaces. Reordering these modules ensures that the internal logic of each interface remains intact, preventing the functional inconsistency common in earlier approaches.
- Machine Learning-Accelerated Model Checking: Exhaustive testing of every possible redesign is computationally prohibitive. OACAL uses a hybrid approach: it performs traditional model checking on a small subset of designs to train a Gradient Boosted Trees (GBT) model. This model then acts as a high-speed filter, predicting which other designs are likely to satisfy security requirements.
- Requirement Degradation: In cases where no design meets every requirement, OACAL allows developers to prioritize and "degrade" or waive lower-priority requirements to find the best possible suboptimal solution.
Key Results and Insights
The framework was evaluated using ticket-booking and online gifting systems, as well as synthetic benchmarks. Key findings include:
- Superior Coverage: In an online gifting case study, OACAL identified 46 valid revisions that satisfied all requirements. In contrast, the previous state-of-the-art tool found only one, which was practically unusable as it required users to pay after logging out.
- Massive Speedup: OACAL's efficiency scales significantly better than previous methods. For a system with 9 modules, OACAL was approximately 10 times faster in the average case and 20 times faster in the worst-case scenario compared to the OASIS baseline.
- High Predictive Accuracy: The machine learning component achieved an F1-score between 0.96 and 1.0, demonstrating that formal model checking results are highly learnable and that ML can effectively serve as a proxy for rigorous formal methods in large search spaces.
Conclusion
OACAL demonstrates that preserving the internal logic of UI components while optimizing their sequence is key to maintaining user-friendly, secure software. By treating model checking as a learning task, the framework overcomes the redundancy traps of previous methods, offering a scalable solution for modern software architecture security.
Summary updated