AI solutions are increasingly being deployed for real-world applications, to support human decision-making or even to make automated decisions. However, research has shown that AI models are trained highly specific to the datasets and not perform as well on novel unseen inputs. More precisely, AI models may fail when inputs are subject to shifts in the data distributions, for instance, due to adversarial attacks (such as adversarial examples and data poisoning) or changes in the deployment environment. These failures can sometimes cause catastrophic consequences, especially in safety-critical applications like self-driving vehicles and automated disease diagnosis. We therefore aim to develop AI models that are robust to such shifts (safe) and failing which, recover from these shifts (resilient).

In addition, Deepfake technology fabricates realistic images, audio, and videos, making it difficult for AI systems to distinguish between genuine and fake content, increasing the complexity of detection and defense. The misuse of Deepfakes could mislead or manipulate AI systems, compromising their decision-making safety and reliability.  Moreover, the privacy and ethical issues posed by Deepfakes challenge the fairness and transparency of AI, highlighting the need for more robust safeguards and detection mechanisms to ensure the security of these systems. 

Ensuring the safety of an AI system involves understanding how it works and the decision it makes. This allows us to then build in appropriate fail-safe mechanisms and fix issues when they arise. Having an interpretable and explainable AI system would help build trust in its predictions.