The development of advanced machine learning techniques has led to the emergence of deepfakes, very realistic AI-generated synthetic images. Deepfakes create an issue of trust regarding digital media and image authenticity. This paper aims to evaluate ResNet and DenseNet models feature extraction methods in detecting deepfake images. The experiment dataset comprises 60,000 images evenly divided into 30,000 real images and 30,000 AI-generated images. Transfer learning was implemented by both models with pre-trained ImageNet weights. Accuracy, precision, recall, and F1 score also used to measure the models performance. According to the findings, ResNet is superior to DenseNet with 98.33% accuracy. Nevertheless, the recall of DenseNet reaches 99.34%, showing better sensitivity than ResNet despite the fact that it has lower accuracy of 95.42%. The Grad-CAM visualization proves that there are distinctive feature-focusing tendencies. ResNet tends to focus on global features, whereas DenseNet distributes its attention among local features.
The use of Federated Artificial Intelligence on Cloud Infrastructure is creating a new urgency for Privacy Preserving Security Mechanisms for Sensitive Data and for preserving Computation Efficiency and Model Utility. Traditional methods for achieving Privacy Preservation use Static Homomorphic Encryption Configurations and Uniform Trust Assumptions, and therefore do not take into consideration the contextually dependent differences in the sensitivity of the data, the reliability of the client and the adversarial risks associated with each data element. Most existing solutions are subject to Excessive Computational Overhead; Model Accuracy Degradation; Inference Attacks; Poisoning Attacks; and treatment of Encrypted Information as Context Agressive Entities. Therefore, these limitations prevent the Practical Adoption of Secure Federated Learning in High-Stake Domains such as Healthcare; Finance; Critical Infrastructure in Process; etc. This research presents CASE-HEFL, a Context Adaptable Semantic Homomorphic Encryption Framework that enables the creation of Secure, Efficient, and Robust Federated AI Systems in Cloud Environments. The CASE-HEFL framework consists of five analytically grounded modules operating in a closed-loop feedback pipeline. First, Context Adaptable Semantic Encryption Profiling (CASE-P) creates a multi-dimensional sensitivity vector from Local Data; Gradients; Metadata; and Deep Semantic Encoding, providing the ability to precisely quantify the risk associated with each piece of data. Second, Sensitivity Weighted Homomorphic Parameter Synthesis (SW-HPS), creates dynamically encrypted parameter values based upon the Sensitivity Profile of the client, thereby optimizing noise budgets and computational depth, while maintaining Cryptographic Strength. Third, Context Aware Encrypted Gradient Transformation (CA-EGT), applies Stochastic Gradient Shaping to the gradients before encrypting them, thus preventing adversaries from gaining knowledge of the Update Patterns. Fourth, Secure Context Weighted Homomorphic Aggregation (SCW-HA), applies Trust Aware Weights to Client Contributions after they have been encrypted, thereby improving the Resilience of the System against Malicious or Low-Quality Updates without Revealing Plaintext Information Sets. Fifth, Adaptive Secure Model Reconstruction with Contextual Feedback (ASMRCF), Decrypts the Aggregated Updates and Produces Risk Aware Feedback Signals that Refine Subsequent Protection Strategies, allowing the system to continuously adapt to changes during Training Rounds. Compared to traditional Static HE-Based Federated Systems, the Proposed Framework Achieves Significant Improvements in both Privacy Preservation and Computational Efficiency, as well as Robustness. Analytical Evaluations demonstrate improvements in Encryption Overhead of greater than 35%; Enhanced Resistance to Inference and Poisoning Attacks; and Faster Convergence Rates without Sacrificing Model Accuracy. By transforming Homomorphic Encryption from a static Protective Layer into a Context Responsive Security Mechanism, this Research Advances the Feasibility of Trustworthy Federated AI for Sensitive Cloud Applications, and Establishes the Foundation for Next Generation Adaptive Cryptographic Learning Systems.
In this paper, a multimodal fusion framework is proposed that will help fuse diverse agricultural information to facilitate crop monitoring and disease diagnosis. Current methods of agricultural monitoring do not tend to bring together the various types of data, including multispectral satellite imagery, high-resolution UAV thermal data, and real-time ground sensor networks. We suggest that we deal with these restrictions by having a three-stage fusion architecture. Initially, a Cross-Modal Attention Fusion Network (CMAFN) is based on the scaled dot-product attention to train complementary representations between satellite and ground sensor-based data, and allows varying feature interaction. Second, by combining UAV RGB and thermal images, a Hierarchical Spatial-Spectral Fusion Module combines them with Vision Transformer encoders and multi-level U-Net architectures to attain accurate crop segmentation. Third, decision-level fusion mechanism is used to integrate multimodal outputs to enhance predictive performance. It has been experimentally validated on paddy fields in Konkan region with high improvements attaining 96.2% crop classified, with the yield prediction R2 of 0.92 and an improved F1-score of 93.7% in multi-class disease detection with a false positive reduced by 21%. The given strategy enhances as well features correlation between satellite NDVI and soil moisture by 14.7%. Such findings underscore the efficacy of multimodal fusion through cross-attention to use in precision agriculture on a data basis.
The rapid increase in the use of deep learning, despite improving diagnostic accuracy in medical image analysis, has also raised significant concerns regarding regulatory compliance, data protection, and ensuring the privacy of the distribution of patient information. Traditional centralized learning methods are needed combining sensitive medical photographs from various healthcare facilities integrate institutions into a single repository, increasing the possibility that data violations and breaches of data protection legislation such as the GDPR as well as HIPAA. This research proposes a federated deep learning method. A system for categorizing medical images while ensuring privacy, enabling joint model training in decentralized medical centers while keeping actual patient information confidential. The framework employs federated learning to train convolutional neural networks models by iteratively adding updates that have been trained locally, maintain a high classification while guaranteeing the secrecy of the data performance. The tests reveal that the suggested method achieves competitive accuracy and precision, recovery and F1 score compared to centralized models, while significantly increasing regulatory compliance, security, and privacy. The suggested system provides a pragmatic and adaptable method for secure and decentralized evaluation of medical images in real healthcare environments.
Phishing continues to be one of the most prevalent cyber attacks which targets users through misleading emails, malicious URLs and fraudulent websites. The conventional rule-based and black-list-based detection techniques struggle with identifying the changing phishing methods and there is a need to use intelligent detection techniques. In this paper, we perform a thorough survey on the use of machine learning and deep learning methods for phishing detection with a particular focus on the workflow of modern AI-driven cybersecurity framework. In the reviewed architecture, the steps include collection of data from various sources such as phishing URL databases, e-mail dataset and webpage features. Subsequent to this is the process of pre-processing and feature extraction. In the framework, machine learning techniques such as support vector machine, random forest and logistic regression along with deep learning techniques like convolutional neural network and long-short term memory network are employed. This survey will also explore the use of explainable artificial intelligence techniques for achieving more transparency in the cybersecurity application. Comparisons among the existing works show how these approaches differ in terms of accuracy, computation complexity and interpretability. It is observed from the analysis that the hybrid architectures based on machine learning, deep learning and explainable artificial intelligence provide promising ways forward in building phishing detection systems.
