Click For Photo: https://3c1703fe8d.site.internapcdn.net/newman/gfx/news/2019/certifyingat.jpg
When shopping for a watch, you may notice its water resistance rating, which indicates that the watch is warranted to be waterproof to a certain level. What about your neural network? Can one ensure a neural network is "attack proof", meaning that its functionality is robust against adversarial perturbations? If so, how can this be quantified with an attack resistance number? At AAAI 2019, our group of researchers from MIT and IBM Research proposes an efficient and effective method for certifying attack resistance of convolutional neural networks to given input data. This paper is selected for oral presentation at AAAI 2019 (January 30, 2:00-3:30 pm @ coral 1).
Current deep neural network models are known to be vulnerable to adversarial perturbations. A carefully crafted yet small perturbation to input data could easily manipulate the prediction of the model output, including machine learning tasks such as object recognition, speech translation, image captioning, and text classification, to name a few. A lack of robustness to adversarial perturbations incurs new challenges in AI research and may impede our trust in AI systems.
Network - Threat - Model - Attack - Strength
Given a neural network and considering an adversarial threat model in which the attack strength is characterized by the Lp norm of the perturbation, for any data input, its adversarial robustness can be quantified as the minimal attack strength required to alter the model prediction (see Figure 1 in the previous post for a visual illustration). Here, an attack-proof robustness certificate for an input specifies an attack strength ε and offers the following...
Wake Up To Breaking News!
The beatings will continue until moral improves.