5 sonuçlar
Arama Sonuçları
Listeleniyor 1 - 5 / 5
Yayın Unreasonable effectiveness of last hidden layer activations for adversarial robustness(Institute of Electrical and Electronics Engineers Inc., 2022) Tuna, Ömer Faruk; Çatak, Ferhat Özgür; Eskil, Mustafa TanerIn standard Deep Neural Network (DNN) based classifiers, the general convention is to omit the activation function in the last (output) layer and directly apply the softmax function on the logits to get the probability scores of each class. In this type of architectures, the loss value of the classifier against any output class is directly proportional to the difference between the final probability score and the label value of the associated class. Standard White-box adversarial evasion attacks, whether targeted or untargeted, mainly try to exploit the gradient of the model loss function to craft adversarial samples and fool the model. In this study, we show both mathematically and experimentally that using some widely known activation functions in the output layer of the model with high temperature values has the effect of zeroing out the gradients for both targeted and untargeted attack cases, preventing attackers from exploiting the model's loss function to craft adversarial samples. We've experimentally verified the efficacy of our approach on MNIST (Digit), CIFAR10 datasets. Detailed experiments confirmed that our approach substantially improves robustness against gradient-based targeted and untargeted attack threats. And, we showed that the increased non-linearity at the output layer has some ad-ditional benefits against some other attack methods like Deepfool attack.Yayın TENET: a new hybrid network architecture for adversarial defense(Springer Science and Business Media Deutschland GmbH, 2023-08) Tuna, Ömer Faruk; Çatak, Ferhat Özgür; Eskil, Mustafa TanerDeep neural network (DNN) models are widely renowned for their resistance to random perturbations. However, researchers have found out that these models are indeed extremely vulnerable to deliberately crafted and seemingly imperceptible perturbations of the input, referred to as adversarial examples. Adversarial attacks have the potential to substantially compromise the security of DNN-powered systems and posing high risks especially in the areas where security is a top priority. Numerous studies have been conducted in recent years to defend against these attacks and to develop more robust architectures resistant to adversarial threats. In this study, we propose a new architecture and enhance a recently proposed technique by which we can restore adversarial samples back to their original class manifold. We leverage the use of several uncertainty metrics obtained from Monte Carlo dropout (MC Dropout) estimates of the model together with the model’s own loss function and combine them with the use of defensive distillation technique to defend against these attacks. We have experimentally evaluated and verified the efficacy of our approach on MNIST (Digit), MNIST (Fashion) and CIFAR10 datasets. In our experiments, we showed that our proposed method reduces the attack’s success rate lower than 5% without compromising clean accuracy.Yayın Machine learning-based model categorization using textual and structural features(Springer Science and Business Media Deutschland GmbH, 2022-09-08) Khalilipour, Alireza; Bozyiğit, Fatma; Utku, Can; Challenger, MoharramModel Driven Engineering (MDE), where models are the core elements in the entire life cycle from the specification to maintenance phases, is one of the promising techniques to provide abstraction and automation. However, model management is another challenging issue due to the increasing number of models, their size, and their structural complexity. So that the available models should be organized by modelers to be reused and overcome the development of the new and more complex models with less cost and effort. In this direction, many studies are conducted to categorize models automatically. However, most of the studies focus either on the textual data or structural information in the intelligent model management, leading to less precision in the model management activities. Therefore, we utilized a model classification using baseline machine learning approaches on a dataset including 555 Ecore metamodels through hybrid feature vectors including both textual and structural information. In the proposed approach, first, the textual information of each model has been summarized in its elements through text processing as well as the ontology of synonyms within a specific domain. Then, the performances of machine learning classifiers were observed on two different variants of the datasets. The first variant includes only textual features (represented both in TF-IDF and word2vec representations), whereas the second variant consists of the determined structural features and textual features. It was finally concluded that each experimented machine learning algorithm gave more successful prediction performance on the variant containing structural features. The presented model yields promising results for the model classification task with a classification accuracy of 89.16%.Yayın Efficient estimation of Sigmoid and Tanh activation functions for homomorphically encrypted data using Artificial Neural Networks(Institute of Electrical and Electronics Engineers Inc., 2024) Harb, Mhd Raja Abou; Çeliktaş, BarışThis paper presents a novel approach to estimating Sigmoid and Tanh activation functions using Artificial Neural Networks (ANN) optimized for homomorphic encryption. The proposed method is compared against second-degree polynomial and Piecewise Linear approximations, demonstrating a minor loss in accuracy while maintaining computational efficiency. Our results suggest that the ANN-based estimator is a viable alternative for secure machine learning models requiring privacypreserving computation.Yayın ANN activation function estimators for homomorphic encrypted inference(Institute of Electrical and Electronics Engineers Inc., 2025-06-13) Harb, Mhd Raja Abou; Çeliktaş, BarışHomomorphic Encryption (HE) enables secure computations on encrypted data, facilitating machine learning inference in sensitive environments such as healthcare and finance. However, efficiently handling non-linear activation functions, specifically Sigmoid and Tanh, remains a significant computational challenge for encrypted inference using Artificial Neural Networks (ANNs). This study introduces a lightweight, ANN-based estimator designed to accurately approximate activation functions under homomorphic encryption. Unlike traditional polynomial and piecewise linear approximations, the proposed ANN estimators achieve superior accuracy with lower computational overhead associated with bootstrapping or high-degree polynomial techniques. These estimators are trained on plaintext data and seamlessly integrated into encrypted inference pipelines, significantly outperforming conventional methods. Experimental evaluations demonstrate notable improvements, with ANN estimators enhancing accuracy by approximately 2% for Sigmoid and up to 73% for Tanh functions, improving F1-scores by approximately 2% for Sigmoid and up to 88% for Tanh, and markedly reducing Mean Square Error (MSE) by up to 96% compared to polynomial approximations. The ANN estimator achieves an accuracy of 97.70% and an AUC of 0.9997 when integrated into a CNN architecture on the MNIST dataset, and an accuracy of 85.25% with an AUC of 0.9459 on the UCI Heart Disease dataset during ciphertext inference. These results underscore the estimator’s practical effectiveness and computational feasibility, making it suitable for secure and efficient ANN inference in encrypted environments.
