3 ms·
Yes, basically. Imagine in financial markets if you trained a system to identify trends in currency pairs (for example). Generally they would be benign and us
by blunte 5y ago
Yes, basically.
Imagine in financial markets if you trained a system to identify trends in currency pairs (for example). Generally they would be benign and useful, but if the a certain input were present, they would advise an action which the designers would expect and profit from. The users of the system would follow the advice as normal and probably not even know that someone essentially had inside information.
- kiwih 5y agoThere is a name for this in machine learning, it's called a 'poisoning attack'. Essentially you add false data into a training set. An example: https://arxiv.org/abs/1703.01340 https://arxiv.org/abs/1703.01340