Federated Learning (FL) enables collaborative model training without sharing raw data, making it particularly suitable for smart agriculture, where data is sensitive and distributed across multiple stakeholders. However, FL introduces critical security vulnerabilities that can degrade model performance and compromise data confidentiality. This study investigates these vulnerabilities using real-world microclimate time series data from four greenhouses, representing distinct agricultural actors collaborating while retaining data locality. We study multiple adversarial attacks, including Data Poisoning, Backdoor, Model Update Poisoning, Property Inference, and Free Rider attacks, evaluating their impact on predictive performance and model reliability. To mitigate these threats, we evaluate multiple defense strategies, including Differential Privacy, Robust Aggregation, Quantization, Sparsification, Gradient Clipping, and Free Rider Detection mechanisms deployed at both client and server levels. Building on these existing mechanisms, we design a unified multi-layered defense pipeline that integrates data sanitization, adaptive clipping, Trust-Krum aggregation, and privacy-preserving noise injection into a single system-level framework. Experimental results demonstrate that this configuration achieves superior F1-Score across all attack scenarios while maintaining convergence stability. These findings underscore the necessity of integrated, synergistic defenses to ensure the secure and robust deployment of FL in adversarial environments.
Human-in-the-loop (HITL) reinforcement learning can improve exploration and policy acquisition by incorporating external guidance during training. In practice, however, guidance is often noisy, intermittent, and constrained by human effort, especially when intent is inferred from brain–computer interface (BCI) signals. This motivates a central question for reliable HITL-RL: how guidance should be used and when it should be discontinued. This paper presents NeuroRL, a simulation-based HITL-RL framework for studying imperfect BCI-derived intent guidance under controllable reliability. Electroencephalographic (EEG) classifier accuracies estimated from 28 participants parameterize probabilistic guidance, injected as corrective actions in GridWorld and preference-based feedback in Meta-World. We also introduce a warm-up-gated HITL stopping method that calibrates the transition from guided exploration to autonomous learning and activates post-handoff deterioration detection only after this warm-up phase, using Sustained Decline and Peak-Relative Decay signals. Experiments show that guidance reliability and intervention management jointly shape performance: GridWorld remains relatively robust under moderate noise, whereas Meta-World requires higher reliability for stable manipulation learning. Fixed-stop, dense-horizon, cross-accuracy, and intermediate-accuracy analyses show that useful stopping regions shift with guidance reliability and that prolonged imperfect guidance can introduce cumulative bias. Transfer and ablation studies indicate that the prediction-verification protocol identifies useful local stopping regions beyond the primary setting, and that the warm-up gate prevents premature deterioration detection during early exploration. These results support NeuroRL as a controlled framework for analyzing imperfect HITL guidance and a practical intervention management policy for handoff to autonomous reinforcement learning.
While deepfake detection methods have seen significant progress in recent years, current approaches mostly focus on detecting fully synthetic images or partially manipulated images separately. Since in the real world, deepfakes can originate from any such paradigm, it is imperative to develop unified models for detecting both types of deepfakes simultaneously. However, training such models from scratch requires large amounts of annotated data and is computationally limiting. To address this, in this work, we propose a data-efficient adaptation framework, which repurposes a pretrained fully synthetic deepfake detection model to effectively learn partially manipulated image forgeries using a small number of training examples, while maintaining competitive performance on the original synthetic image detection task. We introduce three novel modules: (a) Noise-aware Patch Enhancement (NPE), which captures subtle manipulation artifacts present in partially manipulated images, (b) Adaptive Score Aggregation (ASA), which modulates the influence of the global image-level semantics and the local patch-level artifacts based on the input image, and (c) Multi-Scale Alignment, which enhances both image and patch-level discriminative learning. The proposed modules are generalizable and can be integrated into various existing base architectures. Extensive experiments on 14 datasets spanning both synthetic and partially manipulated image detection tasks demonstrate that our proposed approach achieves the best overall generalization across both settings. https://github.com/debarshigit/DELTA
The current learning-based multi-robot navigation methods face challenges such as low learning efficiency and deployment issues in the real world, limiting their practical applications in industrial scenarios. This paper presents a novel goal-guided dual-policy deep reinforcement learning (DRL) framework for multiple mobile robot navigation. Firstly, a strategically chosen subgoal is positioned along the global reference path, gradually guiding the robot’s progression toward the final goal, thereby significantly improving learning efficiency. Secondly, a dual-policy architecture in the DRL model is proposed: the internal policy is dedicated to optimizing the individual robot’s speed for efficient navigation, while the external policy orchestrates collaborations among multiple robots, defining a comprehensive and effective navigation strategy. The proposed method is implemented and tested in the Robot Operating System (ROS) within diverse scenarios. Experimental results demonstrate the proposed method exhibits high learning efficiency and achieves superior navigation performance, as indicated by reduced travel time and high success rate, surpassing other baseline approaches. Moreover, the proposed method exhibits robust generalization capabilities, which can be migrated to entirely distinct scenarios and deployed on physical robots, thereby providing a promising learning-based solution for the multi-robot navigation problem.
