Security Risks of Continual Learning in AI systems
Continual Learning is a threat vector that attackers can take advantage.
As human beings, we are constantly bombarded with data from all sides.
Our senses work overtime to interpret and make us understand what we see, hear, feel, etc.
Machine Learning copies this human ability through continual learning (CL), in which a model is deployed and continually learns from data in a production setting.
Just like the human brain, the model attempts to understand this continuous stream of data and rapidly adapts and re-learns in a Continual learning mode.
Given how much data is generated every minute, continual learning is necessary to keep models re-trained and updated.
The risks of Continual Learning
Machine Learning models are usually checked quite thoroughly when in development and during deployment time.
The data on which they are trained is checked to be of the highest quality so that their decision-making ability is optimal.
However, what happens when the model goes into production and starts the continual learning process?
What if attackers were able to…
