Research Question: Is Deep Learning secure for Robots?
Han Wu @wuhanstudio
Ph.D. Student at the University of Exeter, in the U.K.
Is Deep Learning secure for Robots?
Intelligent Robots: Deep Learning in Robotics
Adversarial attacks against image classification
Deep Learning Models are differentiable
Deep Learning Models are differentiable
Deep Learning Models are differentiable
One Targeted Attack
Multi Targeted Attack
Multi Untargeted Attack
$$J_1(x, \delta, y_h) = max(\ \sigma(c) * \sigma(p_0) \ )$$
# One Targeted Attack
loss = K.max(K.sigmoid(K.reshape(out, (-1, 8))[:, 4]) * K.sigmoid(K.reshape(out, (-1, 8))[:, 5]))
grads = K.gradients(loss, self.model.input)
$$J_1(x, \delta, y_h) = max(\ \sigma(c) * \sigma(p_0) \ )$$
$$J_2(x, \delta, y_h) = \sigma(c) * \sigma(p_0)$$
# Multi Targeted Attack
loss = K.sigmoid(K.reshape(out, (-1, 8))[:, 4]) * K.sigmoid(K.reshape(out, (-1, 8))[:, 5])
grads = K.gradients(loss, self.model.input)
$$J_1(x, \delta, y_h) = max(\ \sigma(c) * \sigma(p_0) \ )$$
$$J_2(x, \delta, y_h) = \sigma(c) * \sigma(p_0)$$
$$J_3(x, \delta, y_h) = \sigma(c) * \sum\sigma(p_i)$$
# Multi Untargeted Attack
loss = K.sigmoid(K.reshape(out, (-1, 8))[:, 4]) * K.sum(K.sigmoid(K.reshape(out, (-1, 8))[:, 5:]))
grads = K.gradients(loss, self.model.input)