Botnets are a major concern for IoT devices deployed on a large scale. Botnets not only can discover new vulnerable devices but are also proficient in injecting malware into the target devices. The estimated traffic generated by such botnets is of the order of Tbps and the attacks cost billions of dollars to the global economy every year. This dissertation investigates the problem of detecting botnet attacks on IoT devices and proposes a novel, complete botnet detection framework that makes use of concepts in the immune system. The proposed framework consists of ``intelligent'' agents that can ``learn'' new patterns with minimal or no supervision. It provides innate and adaptive detection on the principles similar to the human immune response and performs botnet detection at different levels known as lines of defence. Each line of defence is composed of one or more detection agents that are composed of different types of artificial neural networks. The innate detection mechanism is responsible to detect botnets that are not known yet and provides a general detection and works on the principles of biological cell division process. The adaptive detection mechanism makes use of accurate and low-latency neural networks that are trained for specific behaviours. The core of adaptive detection is composed of a novel system inspired by Hematopoietic Stem cells. Experiments are conducted by making use of IoT traffic data from commercially available IoT devices infected with the most popular open-source botnets. Evaluation results demonstrate the power and capabilities of this framework for efficient botnet detection producing high accuracy, low latency and low false detection rate. This dissertation makes three major contributions to the subject domain including (a) a comprehensive botnet detection framework that can be extended towards general anomalies; (b) an efficient unsupervised binary classification detector for botnets that can be extended towards multiclass classification and (c) an automated neural network structure selection mechanism that removes the need to devise the number of hidden layers and the contained neurons and can guide the learning process without specifying the number of epochs and is able to automatically prune the network as needed.