Figure 7.
The diagram contains Attack Generation and Model Training stages. Input samples, including x 1, receive weak transformation t w and strong transformation t s. The weakly transformed samples pass through f theta to produce anchor features. The strongly transformed samples pass through f theta to produce class features z cls. A dot product combines the anchor and class features. Softmax produces output logits y hat, which are compared with targets y tilde using cross-entropy loss L C E. Delta i is calculated using epsilon sign of the gradient of L C E with respect to t w of x i. Delta 1 is added to t w of x 1 to create an attacked weakly transformed sample. These attacked samples pass through f theta to produce another set of anchor features. A dot product combines these anchor features with the class features. Softmax produces another set of output logits y hat, which are compared with the same targets using L C E. The gradient of L C E with respect to theta is used to update f theta.A new self-supervised adversarial training is developed by enhancing the instance discrimination via two-view augmented data and separate anchor and class samples