Early-exit models improve efficiency by allowing inputs to terminate inference at shallow layers, while class-aware variants further prioritize critical classes for faster processing. Yet existing solutions assume centralized training and cannot accommodate distributed settings where devices may have distinct class priorities. We introduce CAFED (Class-Aware Federated Early-Exit Network), the first framework that integrates class-aware early exits into federated learning. CAFED combines clustered aggregation with layer freezing: shared lower layers capture general features, while personalized upper layers host exits aligned with local priorities. Experiments show that CAFED accelerates inference for prioritized classes by up to 40% and improves accuracy by up to 30% over conventional federated and early-exit baselines, establishing a new path for privacy-preserving collaborative learning with class prioritization.

As the integration of edge and IoT devices continues to surge, the need for streamlined machine learning solutions, notably Deep Neural Networks (DNNs), becomes paramount. However, the inherent computational demands and significant memory requirements of DNNs present hurdles for seamless edge integration. Early-Exit Networks (EENs) have emerged as a solution by adding early exits to DNNs, allowing for early inference and facilitating dynamic deployment on devices with varying capabilities. Despite their advantages, current EEN models uniformly handle samples from all classes, which is suboptimal in many edge scenarios where different classes require varying processing speeds and response times. Our research introduces RACENet, a novel architecture with adaptive early-exit capabilities that adjusts class prioritization in real-time. This flexibility enables RACENet to respond to evolving class priorities during runtime. Extensive evaluations on vision tasks, and network intrusion detection demonstrate that RACENet maintains accuracy while dynamically managing class priorities and accelerating inference for high-priority classes, all without adding additional computational burdens on constrained edge devices.

Edge-based and IoT devices have seen phenomenal growth in recent years, driven by the surge in demand for emerging applications that leverage machine learning models, such as Deep Neural Networks (DNNs). However, a primary drawback of DNNs is their substantial storage/memory needs and high computational overhead, making their adoption in edge devices challenging. This limitation prompted the development of early-exit models like BranchyNet, which enable decisions to be made at earlier stages by incorporating dedicated exits within the architecture’s inner layers. Nonetheless, these existing early-exit models lack control over the specific class that should exit and when. The necessity for such class-aware models is evident in numerous edge applications, where particular high-priority classes must be detected earlier due to their time-sensitive nature. In this paper, we introduce ClassyNet, the first early-exit architecture designed to return only selected classes at each exit. This feature facilitates faster inference times for critical classes, allowing the initial layers to operate on edge devices. This strategy conserves considerable computational time and resources on the edge without compromising accuracy. Through extensive experiments, we show the effectiveness of ClassyNet compared to other models under various scenarios.

The exponential growth of IoT and edge devices has led to their widespread use across various applications. However, the security of these devices remains a significant concern due to their vulnerability to a broad spectrum of cyber-attacks. Network Intrusion Detection Systems (NIDS) are crucial for identifying and mitigating such threats. Traditional NIDS approaches, while effective, struggle to detect sophisticated modern attacks and often require substantial computational power and memory, which may not be feasible for edge devices. Machine learning and neural network-based methods have demonstrated promising improvements in NIDS detection accuracy. Yet, their deployment on resource-constrained edge devices presents a challenge. This has led to the development of Dynamic Neural Networks, an approach that allows models to adapt according to the input, making them more efficient and lightweight. However, these networks are class-agnostic, rendering them unsuitable for handling cases with uneven classification priorities. In this paper, we introduce ClassyNet, a platform designed for efficient, classaware NIDS on edge devices. ClassyNet leverages class-specific feature extraction and a class-specific neural network architecture to enhance intrusion detection efficiency. Experimental results indicate that our proposed approach matches the detection accuracy of traditional machine learning and neural network-based methods while significantly improving resource efficiency.

Edge-based IoT devices have experienced phenomenal growth in recent years due to rapidly increasing demand in various emerging applications which typically utilize machine learning (ML) models such as Deep Neural Network (DNN) and demand low latency and low power consumption. To support the edge requirements, ML models have to support faster inference and less computation. Dynamic DNNs (D2NN) have been proposed to support low-latency and power-saving on edge devices by enabling conditional computations and context dependant activation of the network model for inference; saving computational time and edge resources, hence they are becoming popular for edge applications. In this paper, we show that D2NN are vulnerable to our novel adversarial attack, Dynamic DNN Adversarial attacks (DDAS). Unlike conventional adversarial attacks that target classification accuracy, DDAS targets the IoT device resources such as the battery, latency, and so on. We show that our attack is effective under various attack scenarios with a high attack success rate. We also provide a retraining scheme as a countermeasure to DDAS and show its effectiveness.

With the emergence of machine learning technology that supports the development of synthetic models, many new use cases and challenges are emerging in the fields of computer vision and security. The main model behind this technology is Generative Adversarial Networks (GANs), with their ability to model unknown distributions accurately and perform well in generating synthetic data such as images and videos. However, the application of this technology by the networking community has been lacking. Given this motivation, we introduce MirageNet; our vision for a GAN-based synthetic network traffic generation framework, which can automatically create synthetic network models of protocols, applications, and devices. With the potential to build many applications for privacy, security, and network optimization. In this paper, we present MiragePkt; the first component of MirageNet. It is a GAN-based model to synthetically generate network packets. We describe the different challenges, limitations, and solutions for generating synthetic network packets. Finally, we validate and evaluate the performance of our framework with the synthesizing DNS packets.