Robustness in deep learning models for medical diagnostics: security and adversarial challenges towards robust AI applications

Artificial Intelligence Review, Journal Year: 2024, Volume and Issue: 58(1)

Published: Nov. 8, 2024

The current study investigates the robustness of deep learning models for accurate medical diagnosis systems with a specific focus on their ability to maintain performance in presence adversarial or noisy inputs. We examine factors that may influence model reliability, including complexity, training data quality, and hyperparameters; we also security concerns related attacks aim deceive along privacy seek extract sensitive information. Researchers have discussed various defenses these enhance robustness, such as input preprocessing, mechanisms like augmentation uncertainty estimation. Tools packages extend reliability features frameworks TensorFlow PyTorch are being explored evaluated. Existing evaluation metrics additionally This paper concludes by discussing limitations existing literature possible future research directions continue enhancing status this topic, particularly domain, ensuring AI trustworthy, reliable, stable.

Language: Английский

---

Two-Stage Neural Network Optimization for Robust Solar Photovoltaic Forecasting

Electronics, Journal Year: 2024, Volume and Issue: 13(9), P. 1659 - 1659

Published: April 25, 2024

Neural networks (NNs) have shown outstanding performance in solar photovoltaic (PV) power forecasting due to their ability effectively learn unstable environmental variables and complex interactions. However, NNs are limited practical industrial application the energy sector because optimization of model structure or hyperparameters is a time-consuming task. This paper proposes two-stage NN method for robust PV forecasting. First, dataset divided into training test sets. In set, several models with different numbers hidden layers constructed, Optuna applied select optimal hyperparameter values each model. Next, optimized layer used generate estimation prediction fivefold cross-validation on sets, respectively. Finally, random forest values, from set as input predict final power. As result experiments Incheon area, proposed not only easy but also outperforms models. case point, New-Incheon Sonae dataset—one three various locations—the achieved an average mean absolute error (MAE) 149.53 kW root squared (RMSE) 202.00 kW. These figures significantly outperform benchmarks attention mechanism-based deep learning models, scores 169.87 MAE 232.55 RMSE, signaling advance that expected make significant contribution South Korea’s industry.

Language: Английский

---

Evaluating the Impact of Optimizer Hyperparameters on ResNet in Hanacaraka Character Recognition

Preservation Digital Technology & Culture, Journal Year: 2025, Volume and Issue: unknown

Published: Feb. 24, 2025

Abstract This study evaluates the performance of various optimizers on a ResNet-18 based Convolutional Neural Network (CNN) model for task recognizing Hanacaraka Javanese script characters. The image dataset is divided into three sets: training, validation, and test, with sizes 64 × pixels batch size 64. tested include SGD, Adam, RMSprop, Adagrad, Adadelta, NAdam, Adamax, all learning rate 0.001 trained 10 epochs. results show that NAdam provides best accuracy, precision, recall, F1-Score values reaching 100 %, followed by Adamax metrics above 97 %. Adam Adagrad also demonstrate high metric Meanwhile, SGD shows fairly good an accuracy 93.72 Adadelta adequate 86.58 RMSprop yields lowest 81.74 Accuracy loss graphs indicate offer balance between training validation performance, while exhibit significant instability. highlights importance selecting appropriate optimizer to achieve optimal in character classification, being choices.

Language: Английский

---

Deep Feature Fusion via Transfer Learning for Multi-Class Network Intrusion Detection

Applied Sciences, Journal Year: 2025, Volume and Issue: 15(9), P. 4851 - 4851

Published: April 27, 2025

With the rapid advancement of network technologies, cyberthreats have become increasingly sophisticated, posing significant challenges to traditional intrusion detection systems. Conventional machine learning and deep approaches frequently experience performance degradation when confronted with imbalanced datasets novel attack vectors. To address these limitations, this study proposes a learning-based framework that employs feature fusion through incremental transfer between source target domains. The proposed architecture integrates convolutional neural networks (CNNs) an attention mechanism extract aggregate salient features, thereby enhancing model’s discriminative capacity normal traffic various categories. Experimental results demonstrate model achieves accuracy 94.21% even trained on only 33% available data, outperforming conventional models. These findings underscore effectiveness strategy via in improving capabilities within dynamic evolving cyberthreat environments.

Language: Английский

---

Load Day-Ahead Automatic Generation Control Reserve Capacity Demand Prediction Based on the Attention-BiLSTM Network Model Optimized by Improved Whale Algorithm

Energies, Journal Year: 2024, Volume and Issue: 17(2), P. 415 - 415

Published: Jan. 15, 2024

Load forecasting is a research hotspot in academia; the context of new power systems, prediction and determination load reserve capacity also important. In order to adapt forms day-ahead automatic generation control (AGC) demand method based on Fourier transform attention mechanism combined with bidirectional long short-term memory neural network model (Attention-BiLSTM) optimized by an improved whale optimization algorithm (IWOA) proposed. Firstly, response time, used refine distinction between various types demand, AGC band calculated using Parseval’s theorem obtain sequence. The maximum mutual information coefficient explore relevant influencing factors sequence concerning data characteristics Then, historical daily sequences features are input into Attention-BiLSTM model, automatically find optimal hyperparameters better results. Finally, arithmetic simulation results show that proposed this paper has best performance upper (0.8810) lower (0.6651) bounds (R2) higher than other models, it smallest mean absolute percentage error (MAPE) root square (RMSE).

Language: Английский

---

A laser ultrasonic intelligent inspection method for metal surface defects based on digital twin model

Measurement, Journal Year: 2024, Volume and Issue: 237, P. 115219 - 115219

Published: July 1, 2024

Language: Английский

---

DGL-STFA: Predicting lithium-ion battery health with dynamic graph learning and spatial-temporal fusion attention

Energy and AI, Journal Year: 2024, Volume and Issue: unknown, P. 100462 - 100462

Published: Dec. 1, 2024

Language: Английский

---

Damage of thin target plates impacted by conical projectiles with varying apex angles in ballistic application: experimental, FEA and ANN integrated approach

Deleted Journal, Journal Year: 2024, Volume and Issue: 6(10)

Published: Sept. 28, 2024

Language: Английский

---