5 - Validation of Artificial Intelligence Technologies (VAIT)

As artificial intelligence (AI) becomes more ubiquitous in society and the DoD enterprise, trust in the results provided by AI is of paramount importance. To that end, R-DEX Systems introduces VAIT (Validation of Artificial Intelligence Technologies), a processing suite for assessing AI algorithms for validation, trust in the decision-making process, and robustness against adversarial attack. VAIT performs a number of functions to improve the trust in data interpretation and decision process. These functions include 1) identifying weakness through susceptibility assessment to adversarial attack, 2) strategies for strengthening networks against attack, 3) complementary classifiers to quantify decision reliability and confidence and 4) methods for identifying possible adversarial inputs. Methods incorporated into VAIT include classic white box and black box adversarial attacks, adversarial training, complementary single and cross-domain AI classifiers, and meta-reasoning networks for consistency assessment and adversarial input identification. In Phase I, R-DEX will apply VAIT to open source classifier networks including ResNet, DenseNet, Inception, etc. and proprietary, unclassified network architectures including the R-DEX proprietary Deep Automatic Target Recognition (DATR) suite. In Phase II, will work with the sponsor and strategic partners to assess VAIT against operationally relevant networks and design a real-time VAIT proof of concept for operational assessment.