3 ms·
Or, for even more fun: http://www.cs.columbia.edu/~junfeng/papers/deepxplore-sosp17.pdf http://www.cs.columbia.edu/~junfeng/papers/deepxplore-sosp17... Deep le
by groby_b 9y ago
Or, for even more fun: http://www.cs.columbia.edu/~junfeng/papers/deepxplore-sosp17.pdf http://www.cs.columbia.edu/~junfeng/papers/deepxplore-sosp17...
Deep learning exploits at a button press.
Quoth: For all tested DL models, on average DeepXplore generated one test input demonstrating incorrect behavior within one second while running only on a commodity laptop.
One second for triggering a failure, on any tested model. On a commodity laptop.
That's just the technology you want to entrust with intelligence tasks.
- tormeh 9y agoThere has to be some kind of DNN research on using methods such as these to enhance accuracy of DNNs by using the antagonist examples as input. You could just take your existing labeled data and add light-medium noise to all your data (easiest with images?) and thereby get a lot more training data less prone to overfitting. It's so obvious to me so someone must have thought of it already. So my guess is that these kinds of easy antagonist examples won't be a thing for very long.
- groby_b 9y agoRule of thumb: If you find a blatantly obvious solution and you're not an expert in the field, assume people who are have thought of it and discarded it - or it would exist. (Note: I'm not an expert in the field, either. But the above rule works most of the time :)
- abrichr 9y agoIt's worth pointing out that their technique requires access to the internals of the network: First, we compute the gradient of the outputs of the neurons in both the output and hidden layers with the input value as a variable and the weight parameter as a constant. A well designed production system would not support this.