The gradient attack problem has recently been studied to increase the awareness of people on privacy risks in federated learning. However, this attack is constrained under specific conditions, such as small image batch sizes and low image resolutions. To address this challenge, we introduce a new three-phase image recovery architecture called EncGradInversion, which harnesses the power of image encoding and the shared gradient inversion. In the first phase, we attempt to extract the representation for all of the images using the gradient at the final layer. Then, in the second phase, the extracted encoding of a specific image is leveraged for reconstructing the image by matching the representation of dummy and approximated images. This allows a parallel algorithm to accelerate the image recovery. In the final phase, the reconstructed images are fine tuned using the shared gradient of the whole network. In the second and third phases, we formulate an optimization problem to minimize the discrepancy between the shared and reconstructed gradients, while preserving the smoothness and natural appearance of the reconstructed images. Evaluated on various datasets and deep learning models, EncGradInversion shows its superiority to recover the original training images with resolutions as high as$1024\times 1024$and with the batch size of 512. Furthermore, the proposed architecture outperforms existing counterparts with a factor of up to 9.8 and 6.04, in terms of structural similarity performance and attack time.

To reduce network intrusion detection latency in a high volume of data traffic, on-device detection with neuron pruning has been widely adopted by eliminating ineffective connections from a densely connected neural network. However, neuron pruning has a serious problem called output separation in which some parts of neurons can easily be pruned in the middle and become isolated from the rest of the network. To this end, we introduce a solution called the conservation of output links (COOL) pruning method that iteratively preserves a set of effective connections to avoid neuron isolation. We first evaluate COOL on MNIST and CIFAR-10 data sets as well as programmable networking devices, such as P4-supported switches. The experimental results show that COOL outperforms existing methods in terms of both detection time and classification accuracy, especially in extremely sparse networks. Compared to three representative pruning methods, our COOL-based classification model performs at least 25% more accurately with the upper bound for the pruning probability. To further display the effectiveness of COOL-based intrusion detection, we formulate a novel detection time minimization problem by assigning suitable detection models for switches in Internet of Things (IoT) under performance requirements and resource limitations. The experimental results demonstrate that our COOL algorithm is particularly useful for delay-critical and high-traffic applications.

Due to the outbreak of recent network attacks, it is necessary to develop a robust network intrusion detection system (NIDS) that can quickly and effectively identify the network attack. Although the state-of-the-art detection algorithms have shown quite promising detection performance, they suffer from computationally intensive operations and large memory footprint, making themselves infeasible to applications at the resourceconstrained edge devices. We propose a lightweight yet effective NIDS scheme that incorporates a stacked autoencoder with a network pruning technique. By removing a set of ineffective neurons across layers in the autoencoder network with a certain probability based on their importance, a considerably large portion of relatively nominal training parameters are reduced. Then, the pruned and pretrained encoder network is used as-is and is connected with a separate classifier network for attack type inference, avoiding a full retraining from scratch. Experimental results indicate that our stacked autoencoder-based classification network with probabilistic feature extraction has outperformed the state-of-the-art NIDSs in terms of attack detection rate. Further, we have shown that our lightweight NIDS scheme has significantly reduced the computational complexity throughout the architecture, making it feasible to the edge, while maintaining a similar attack type detection quality compared with its original fully connected neural network.