Resilient Machine Learning System (rMLs)

The recent advances in big data and machine learning (ML) algorithms drive the proliferation of data-driven intelligent applications in many areas including advanced command and control (AC2) missions, decision support systems, recommender systems, computer vision, autonomous vehicles, as well as intrusion detection and prevention systems. Many ML algorithms have been integrated in decision support applications to improve analysis performance and prediction. Due to the wide-spread usage of ML in critical decision processes (e.g., DoD AC2 mission critical applications), there is an exponential growth in cyberattacks that target the ML algorithms and consequently influence their decision process in favoring of the attackers. The adversarial ML (AML) approach has been demonstrated in many domains, including subverting fraud detection, bypassing content filters, or malware detection, misleading autonomous navigation systems and support to target recognition.   In this Phase II project, our goal is to focus on securing the ML system as opposed to using the ML system to build security measures for a wide range of applications in DoD and commercial markets. We will develop a resilient ML ensemble (rMLE) architecture that ensures resilient operations of the ML system from any type of ML attacks. The rMLE is based on Moving Target Defense (MTD), where the rMLE dynamically changes the ML algorithms used in the ML pipeline, thus ensuring resilience to adversarial attacks. In the rMLE approach, even if the attacker succeeds in maliciously impacting one of the ML algorithm, it will be guaranteed to fail against the other ML algorithms used in the ensemble and hence, the rMLE will continue to operate normally in spite of the adversarial learning attacks against the ML ensemble.  The novel features of the proposed rMLE approach are as follows: Data Air Gap Transformation – By separating the data spaces used to train the ML algorithms from the original data spaces, the attackers assume that the ML algorithms were trained using the original data spaces. The DAG transformation will significantly decrease the attacker’s ability to launch successful ML attacks. Random Selection of ML Algorithms – By randomly selecting the ML algorithms to be used during any data analytics cycle, we make it extremely difficult for the attackers to learn from studying the ML system results as well as the ML algorithms being used because the ML algorithms and DAG transformation will be continuously changing in each ensemble. ML Threat Modeling, Analysis and Mitigation Methodology. This will enable us to identify all vulnerabilities that can be exploited by ML attackers, determine their impacts on AC2 missions, Prioritize the mitigation and response methods. rMLE Design and Evaluation Tool – The tool will be used to enable users to experiment with and evaluate different ML algorithms, MTD schemes and DAG transformation methods, and the frequency of the changes in ML algorithms.