A Deep Learning Approach to Web Bot Detection Using Mouse Behavioral Biometrics

Detection of Advanced Web Bots by Combining Web Logs with Mouse Behavioural Biometrics

“…Regarding the detection of web bots based on their behaviour, the most prominent approach is the use of machine learning and, more specifically, the use of classification [20,34,37] or clustering algorithms [3,13,38]. Furthermore, detection algorithms proposed in literature use either web logs [3,12,13,29,34,37,38] or mouse movements [12,42] to detect web bots. Additionally, the detection process can be performed either offline (i.e., decide after the end of a session whether it is from a bot), or online by performing an estimation during the session [12,29].…”

“…Latest research has introduced the use of mouse movements for the detection of web bots [12,42]. Mouse movements can either be represented as images and used directly as input to Convolutional Neural Networks (CNNs) [42] or have several high level actions extracted from them, such as click, point-and-click, and drag-anddrop [12]. For the latter, these actions accompanied with their properties, such as the distance of the mouse move, its duration, and efficiency, are used to train classification models.…”

“…Stevanovic et al [37] web logs simple accuracy>0.99, F-score>0.80 (imbalanced classes) Cabri et al [10] web logs simple F-score>0.955 Iliou et al [20] web logs simple precision, recall, and balanced accuracy >0.95 moderate and advanced balanced accuracy=∼0.55 Chu et al [12] mouse movements between moderate and advanced TPR=0.63∼0.99 (based on the bots' complexity), TNR>0.99 Wei et al [42] mouse movements between moderate and advanced TNR=0.93∼0.99 (based on the bots' complexity)…”

“…The module that uses web logs extracts the most promising features for web bot detection as they have been identified in the literature and creates an ensemble classifier that combines several well-established classifiers that have shown promising results in the web bot detection problem [20]. The second module uses the mouse movements of visitors on each web page to train models that can identify web bot and human mouse movement behaviour [42]. The two modules are combined in a novel way to enable the capturing the different temporal characteristics of the web logs and the mouse movements, as well as the spatial characteristics of the mouse movements.…”

“…Current research on machine learning-based web bot detection uses either web logs [3,12,13,29,34,37,38] or mouse movements [12,42] to detect web bots. Even though the aforementioned techniques are highly accurate, they do not address one key aspect of the problem, which is that web bots might try to exhibit both a browser like fingerprint and a humanlike behaviour to avoid detection [15,20].…”