Deep Reinforcement Learning Driven Applications Testing

Applications have become indispensable in our lives, and ensuring their correctness is now a critical issue. Automatic system test case generation can significantly improve the testing process for these applications, which has recently motivated researchers to work on this problem, defining various approaches. However, most state-of-the-art approaches automatically generate test cases leveraging symbolic execution or random exploration techniques. This led to techniques that lose efficiency when dealing with an increasing number of program constraints and become inapplicable when conditions are too challenging to solve or even to formulate. This Ph.D. thesis proposes addressing current techniques' limitations by exploiting Deep Reinforcement Learning. Deep Reinforcement Learning (Deep RL) is a machine learning technique that does not require a labeled training set as input since the learning process is guided by the positive or negative reward experienced during the tentative execution of a task. Hence, it can be used to dynamically learn how to build a test suite based on the feedback obtained during past successful or unsuccessful attempts. This dissertation presents three novel techniques that exploit this intuition: ARES, RONIN, and IFRIT. Since functional testing and security testing are complementary, this Ph.D. thesis explores both testing techniques using the same approach for test cases generation. ARES is a Deep RL approach for functional testing of Android apps. RONIN addresses the issue of generating exploits for a subset of Android ICC vulnerabilities. Subsequently, to better expose the bugs discovered by previous techniques, this thesis presents IFRIT, a focused testing approach capable of increasing the number of test cases that can reach a specific target (i.e., a precise section or statement of an application) and their diversity. IFRIT has the ultimate goal of exposing faults affecting the given program point.