We present a comprehensive study on applying machine learning to detect distributed Denial of service (DDoS) attacks using large-scale Internet of Things (IoT) systems. While prior works and existing DDoS attacks have largely focused on individual nodes transmitting packets at a high volume, we investigate more sophisticated futuristic attacks that use large numbers of IoT devices and camouflage their attack by having each node transmit at a volume typical of benign traffic. We introduce new correlation-aware architectures that take into account the correlation of traffic across IoT nodes. We extensively analyze the proposed architectures by evaluating five different neural network models trained on a dataset derived from a 4060-node real-world IoT system. We observe that long short-term memory (LSTM) and a transformer-based model, in conjunction with the architectures that use correlation information of the IoT nodes, provide higher performance (in terms of F1 score and binary accuracy) than the other models and architectures, especially when the attacker camouflages itself by following benign traffic distribution on each transmitting node. For instance, by using the LSTM model, the distributed correlation-aware architecture gives 81% F1 score for the attacker that camouflages their attack with benign traffic as compared to 35% for the architecture that does not use correlation information. We validate the effectiveness of our proposed detection mechanism by implementing it on a real testbed. We also investigate the performance of heuristics for selecting a subset of nodes to share their data for correlation-aware architectures to meet resource constraints.

We investigate the application of artificial intelligence to cybersecurity, to contribute to the safe and secure growth of the internet of things (IoT). Specifically, we train and evaluate different neural networks models to detect distributed denial of service (DDoS) attacks in a large-scale IoT system. We consider futuristic attacks launched by sophisticated malicious entities that take over multiple distributed IoT nodes and are able to disguise their intrusion by closely mimicking the benign traffic of the network. Using data from prior work, we find that a truncated Cauchy distribution is a suitable fit for benign traffic volume from IoT devices, and we model the attack traffic volume as following the same distribution but with different parameters for location and scale. We emulate both benign and attack traffic by overlaying these traffic volume distributions on top of an activity status data trace from a real urban IoT deployment consisting of about 4000 nodes. Using our enhanced dataset, we compare four neural network models: multi-layer perceptron (MLP), convolutional neural network (CNN), long short-term memory (LSTM), and autoencoder (AEN), analyzing their performance as a function of a parameter that measures the deviation of the attacks from the benign data. We observe that all four models are sensitive to the distance between benign and attack traffic. We further observe that LSTM gives the best overall performance in terms of both high accuracy and high recall.

As IoT deployments grow in scale for applications such as smart cities, they face increasing cyber-security threats. In particular, as evidenced by the famous Mirai incident and other ongoing threats, large-scale IoT device networks are particularly susceptible to being hijacked and used as botnets to launch distributed denial of service (DDoS) attacks. Real large-scale datasets are needed to train and evaluate the use of machine learning algorithms such as deep neural networks to detect and defend against such DDoS attacks. We present a dataset from an urban IoT deployment of 4060 nodes describing their spatio-temporal activity under benign conditions. We also provide a synthetic DDoS attack generator that injects attack activity into the dataset based on tunable parameters such as number of nodes attacked and duration of attack. We discuss some of the features of the dataset. We also demonstrate the utility of the dataset as well as our synthetic DDoS attack generator by using them for the training and evaluation of a simple multi-label feed-forward neural network that aims to identify which nodes are under attack and when.