Recent successful adversarial attacks on face recognition show that, despite the remarkable progress of face recognition models, they are still far behind human intelligence for perception and recognition. It reveals the vulnerability of deep convolutional neural networks (CNNs) as a state-of-the-art building block for face recognition models against adversarial examples, which can cause certain consequences for secure systems. Gradient-based adversarial attacks have been widely studied and proved successful against face recognition models. However, finding the optimized perturbation per each face needs to submit a significant number of queries to the target model. In this paper, we propose a recursive adversarial attack on face recognition using automatic face warping, which needs an extremely limited number of queries to fool the target model. Instead of a random face warping procedure, the warping functions are applied on specific detected regions of face like eyebrows, nose, lips, etc. We evaluate the robustness of the proposed method in the decision-based black-box attack setting, where the attackers have no access to the model parameters and gradients, but the target model provides hard-label predictions and confidence scores.

In this paper, we use data projection to address uncertainty problem in imbalanced data classification. Instead of resampling the data which causes uncertainty, we project the minority and majority instances into new space using a novel technique called Cross-Concatenation. To project the minority instances into a novel space, we concatenate each minority instance with all majority instances to form M * N new double size data where, M and N are the size of minority and majority classes, respectively. The same procedure is repeated to project majority instances into new space by concatenating each majority instance with all minority instances to form N * M new double size data. Our experiments show that, Cross-Concatenation can provide the classifiers sufficient data to train more efficient models since the projected classes are no longer skewed. After training the model, each test data is concatenated with the centroid of minority and majority classes to create two different instances. Afterwards, the highest probability returned from the trained model is used as a metric to assign the label to it. Our experimental results show that, the proposed method can significantly d ecrease the uncertainty in imbalanced classification with competitive results comparing to the SMOTE and its variants which are the most popular over-sampling techniques in terms of precision, recall, F1 and Area Under Curve (AUC).

In this paper, we propose a new framework called Privacy Preserving Proxy (PPP) to protect the online models. PPP is designed like a gateway to work on behalf of the users when requesting service, potentially masking the true origin of the online model from client. Instead of connecting directly to the machine learning-as-a-service(MLaaS) the clients submit the request to the PPP, which evaluates the request and executes some functions to maximize the safety of the models. PPP can potentially host a wide range of functions or triggers to control the complexity of the request or provide additional benefits such as privacy and security. We propose two different methods including Random Model Assignment and Intelligent Rounding Policy which can function inside PPP to protect the privacy of online models. To test the PPP we test TL-GAN based face reconstruction attack as a sophisticated approach which can reveal inefficiency of rounding policy as countermeasure. Our experiments show that, PPP can potentially nullify such complicated attacks in the future.

Researchers deal with the class imbalanced problem in many real-world applications and GAN based data augmentation is considered as an efficient approach to address this problem. GANs need a huge training data to generate efficient augmented data. However, the required sufficient training data is not available in many research areas. In this paper, we introduce a new concept called virtual big data to address this problem. We prove that, virtual big data can provide the GANs sufficient training data to generate efficient augmented data with less mode collapse and vanishing generator gradients problems. We show that, the curse of dimensionality which is considered as a negative factor in machine learning can play a positive role to solve vanishing generator gradients via making discriminator less perfect. First, we transform the training data from n dimensional space into m dimensional space where, m = c * n and c is concatenation factor. To do so, c different training instances are selected and concatenated to each other to form a c * n dimensional instance. Increasing the dimension of training data from n to c * n is key to increase the number of training instances from N to C(N, c). Transformed training data are called virtual big data since they differ original training instances in terms of size and dimension. Our experiments show that, V-GAN, a GAN trained by virtual big data can outperform standard GANs when it comes to deal with extremely scarce training data. Furthermore, V-GAN can outperform traditional oversampling techniques in terms of precision, F1 score and Area Under Curve (AUC) score.