Search for: All records

Differentially Private Stochastic Gradient Descent (DP-SGD) has become a widely used technique for safeguarding sensitive information in deep learning applications. Unfortunately, DP-SGD’s per-sample gradient clipping and uniform noise addition during training can significantly degrade model utility and fairness. We observe that the latest DP-SGD-Global-Adapt’s average gradient norm is the same throughout the training. Even when it is integrated with the existing linear decay noise multiplier, it has little or no advantage. Moreover, we notice that its upper clipping threshold increases exponentially towards the end of training, potentially impacting the model’s convergence. Other algorithms, DP-PSAC, Auto-S, DP-SGD-Global, and DP-F, have utility and fairness that are similar to or worse than DP-SGD, as demonstrated in experiments. To overcome these problems and improve utility and fairness, we developed the DP-SGD-Global-Adapt-V2-S. It has a step-decay noise multiplier and an upper clipping threshold that is also decayed step-wise. DP-SGD-Global-Adapt-V2-S with a privacy budget of 1 improves accuracy by 0.9795%, 0.6786%, and 4.0130% in MNIST, CIFAR10, and CIFAR100, respectively. It also reduces the privacy cost gap by 89.8332% and 60.5541% in unbalanced MNIST and Thinwall datasets, respectively. Finally, we develop mathematical expressions to compute the privacy budget using truncated concentrated differential privacy (tCDP) for DP-SGD-Global-Adapt-V2-T and DP-SGD-Global-Adapt-V2-S. 

Abstract The most common eye infection in people with diabetes is diabetic retinopathy (DR). It might cause blurred vision or even total blindness. Therefore, it is essential to promote early detection to prevent or alleviate the impact of DR. However, due to the possibility that symptoms may not be noticeable in the early stages of DR, it is difficult for doctors to identify them. Therefore, numerous predictive models based on machine learning (ML) and deep learning (DL) have been developed to determine all stages of DR. However, existing DR classification models cannot classify every DR stage or use a computationally heavy approach. Common metrics such as accuracy, F1 score, precision, recall, and AUC-ROC score are not reliable for assessing DR grading. This is because they do not account for two key factors: the severity of the discrepancy between the assigned and predicted grades and the ordered nature of the DR grading scale.  This research proposes computationally efficient ensemble methods for the classification of DR. These methods leverage pre-trained model weights, reducing training time and resource requirements. In addition, data augmentation techniques are used to address data limitations, improve features, and improve generalization. This combination offers a promising approach for accurate and robust DR grading. In particular, we take advantage of transfer learning using models trained on DR data and employ CLAHE for image enhancement and Gaussian blur for noise reduction. We propose a three-layer classifier that incorporates dropout and ReLU activation. This design aims to minimize overfitting while effectively extracting features and assigning DR grades. We prioritize the Quadratic Weighted Kappa (QWK) metric due to its sensitivity to label discrepancies, which is crucial for an accurate diagnosis of DR. This combined approach achieves state-of-the-art QWK scores (0.901, 0.967 and 0.944) in the Eyepacs, Aptos, and Messidor datasets. 

An efficient fake news detector becomes essential as the accessibility of social media platforms increases rapidly. Previous studies mainly focused on designing the models solely based on individual data sets and might suffer from degradable performance. Therefore, developing a robust model for a combined data set with diverse knowledge becomes crucial. However, designing the model with a combined data set requires extensive training time and sequential workload to obtain optimal performance without having some prior knowledge about the model's parameters. The presented study here will help solve these issues by introducing the unified training strategy to have a base structure for the classifier and all hyperparameters from individual models using a pretrained transformer model. The performance of the proposed model is noted using three publicly available data sets, namely ISOT and others from the Kaggle website. The results indicate that the proposed unified training strategy surpassed the existing models, such as Random Forests, convolutional neural networks, and long short-term memory, with 97% accuracy and achieved the F1 score of 0.97. Furthermore, there was a significant reduction in training time by almost 1.5 to 1.8 × by removing words lower than three letters from the input samples. We also did extensive performance analysis by varying the number of encoder blocks to build compact models and trained on the combined data set. We justify that reducing encoder blocks resulted in lower performance from the obtained results.