A Hybrid MAML-reptile Based Few-shot Learning Approach for Securing Fog-iot Networks against Maleficent Behaviors

Security in Fog based IoT networks has a major problem where some IoT devices may be compromised due to attacks. This creates vulnerability to the sensitive data flow between Fog and IoT devices. Also, due to heterogeneity in network structure, the probability of attacks in the network is more. To analyze this dynamic and complex structure of fog networks, few shot learning is be a good solution to learn patterns more accurately and identify the maleficent or normal behavior of a device. In deep learning, few shot learning technique is a technique that uses less amount of labeled data for data processing that is more efficient than the traditional deep learning models. In this work, a hybrid MAML (Model Agnostic Meta Learning)-Reptile based few shot learning approach is proposed that secures the fog based IoT infrastructure from variety of attacks by detecting the attacks more accurately. The few shot models considered are Prototypical Networks, Matching Networks, MAML, and Reptile for selection of best model to be run at the fog server for attack detection. The fog node uses the best model to detect the attacks and broadcasts the local behavior list (LBL) to cloud and other fog nodes to generate the global behavior list (GBL) for sharing more attack information to the IoT device layer. Here, we consider standard datasets such as UNSW-NB-15, NSL-KDD, and CICIDS for performing implementations. The performance of the models is analyzed using accuracy, recall, f1-score, precision, inference time, training time, AUC-ROC, cost, energy consumption, and processing latency. From the results, it is observed that the proposed MAML-Reptile hybrid model performs better than other standard models in detecting maleficent behaviors more accurately.