4 ms·
The article only describes extremely ideal conditions. Use 'in the wild' is still untested, from what I understand. Also, I don't think they provided any code t
by ColanR 7y ago
The article only describes extremely ideal conditions. Use 'in the wild' is still untested, from what I understand. Also, I don't think they provided any code to generate the disguised version of the adversarial patches.
- awb 7y agoWhat happens once the training data includes these adversarial patches? Wouldn't the AI learn to detect them accurately?
- WilTimSon 7y agoI presume new ones could be created based on the newer generation of surveillance tech, which would then become obsolete and the cycle would restart until the tech is sophisticated enough to not be bothered by them. At that point, more complex means would have to be developed.
- AlexCoventry 7y agoYes. In fact, including adversarial examples can be a great regularization technique, which leads to a system learning much more natural representations.