Adversarial Training and Machine Learning

Abstract

Artificial intelligence can be described as the study of machine systems with the ability to reason and perform cognitive functions in a manner almost similar to human intelligence. Artificial Intelligence has grown in prominence over the past few decades. Today, artificially intelligent algorithms control complex banking and financial systems, self-driving cars, and even news feeds. Machine Learning as a subfield has been at the forefront of AI adoption in several industries and sub-fields of AI. Today, ML is used in several applications such as facial recognition, malware detection, robotics, and self-driving cars. Like every computer-based system, however, ML poses its own set of challenges in cybersecurity. This is made harder by the fact that it is increasingly being adopted at a much faster rate than other technological systems. This has great risk not only for businesses and clients who use AI systems but also for the adoption of AI. This paper explored the cyber risks and the potential impact of AI. It detailed the external and internal organizational risks associated with the adoption of AI. In particular, it was concerned with Adversarial Machine Learning as a cybersecurity risk and its potential implications. A review of the literature found several organizations had experienced Adversarial Machine Learning as a threat. A number of these attacks were evasion attacks that manipulated data sets and were therefore hard to detect. This paper used stochastic adversarial training methods to show Adversarial Training can make ANNs adversarial robust. This paper recommends the use of Adversarial Training as a way of combatting Adversarial ML attacks.