Social Media Bot Detection using Dropout-GAN

Various techniques have been proposed for detecting bots on social media platforms, including user behavior analysis, content analysis, and network analysis [20]. One of the well-known works in user behavior analysis is the BotOrNot system [5] which relies on over a thousand features to determine whether a Twitter user’s behavior resembles that of a bot. This work was later expanded upon in [30] where the authors trained the model on a new dataset, and demonstrated which features they leveraged in their technique.

The Random Forest (RF) technique saw early success, as detailed in [4]. Another technique previously used for bot detection is boosting. For example, in [18] a modified version of AdaBoost called BoostOR was used with success, where BoostOR is optimized for F1 score by balancing the precision and recall of the classifier. Convolutional Neural Networks (CNN) have also been applied, with one example of such research being [1].

A clustering-based approach is considered in [12] where a Centroid Initialization Algorithm is used to identify malicious bot users. Some previous work has explored graph-based techniques [33], while others use deep learning models such as Long Short-Term Memory (LSTM) [16] in an attempt to deal with the challenges associated with bot detection. In addition, ensemble techniques have also been proposed, in tandem with appropriate feature selection [28]. Other research efforts have investigated the utilization of Natural Language Processing (NLP) techniques, including word embeddings [22], for improved feature encoding and representation. These prior investigations have shed light on the effectiveness of different methodologies and paved the way for further advancements in the realm of bot detection and analysis. Building upon these existing foundations, this study aims to contribute novel insights by employing an extensible open framework that integrates multiple discriminators in the training process.

Work has been previously done using GANs to identify social media bots [19]. However, in our approach, we propose a simpler architecture, and we expand the concept of using GANs for social media bot detection to a multilingual dataset. The dataset we use includes tweets from 54 languages, and our method also outperforms the previous work in [19], in terms of accuracy and other relevant metrics such as precision, recall, F1 score, and so on.

Bot detection techniques can be broadly classified into two categories: feature-based techniques and graph-based techniques. Feature-based techniques rely on features such as user behavior, content, and so on, whereas graph-based techniques emphasize the use of graph analysis to detect bot activity. Such graph analysis may focus on aspects such as network properties, graph structure, and spatial information like proximity to other humans or bots [20].

Generative Adversarial Networks (GANs) are a class of deep learning models that were introduced in [13]. GANs consist of two components, namely, a generator and a discriminator. The generator produces synthetic data that is meant to resemble real training data, while the discriminator tries to distinguish between real data and the synthetic data produced by the generator.

GANs have been applied to a wide range of tasks, such as image and text generation [15], and anomaly detection [23]. In the context of bot detection on social media platforms, GANs can be used to model the behavior of bots by training the generator to produce synthetic bot behavior that resembles real bot behavior.

The discriminator in the GAN can be used to detect social media bots by training it to distinguish between genuine user behavior and synthetic bot behavior. The discriminator can be trained on a dataset of labeled data where the labels indicate whether a user is a bot or not. The discriminator then learns to identify patterns in the data that are characteristic of bot behavior. Once the discriminator is trained, it can be used to detect bots in new data that it has not seen before.

The generator in the GAN can be used to model social media bots by training it to produce synthetic bot behavior that resembles real bot behavior. The generator is trained on a dataset of labeled data where the labels indicate whether a user is a bot or not. The generator then learns to produce synthetic bot behavior that is similar to the behavior of real bots in the dataset. This synthetic data can serve to strengthen the discriminator, while the generator-produced data provides additional examples of potential bot behavior.

The MGTAB [24] dataset exhibits substantial potential as a foundation for bot research due to its remarkable wealth of data. The authors created a new graph-based dataset by selecting 100 seed accounts actively engaged in discourse about Japan’s plan to release nuclear wastewater into the ocean, then collecting the 10,000 most recent tweets for each user, eventually accumulating a total 1,554,000 users and 135,450,000 tweets. After cleaning, they retain 410,199 users and roughly 40 million tweets. However, not all these users and tweets are labeled; only 10,199 users and their associated tweets were manually annotated to describe whether they were bot accounts or human accounts, and also have other labels that are relevant for stance detection (“neutral”,“against”,“support”), but were not used in our research. The authors of [24] employ many contemporary machine learning techniques, as well as graph-based machine learning techniques to determine their performance on the dataset. These results serve to benchmark the performance against previous datasets as well. By virtue of its versatile combination of feature-based and graph-based characteristics, this dataset is highly conducive to effective experimentation and analysis.

Twibot-22 [9] also stands out as a notable dataset, particularly due to influencing the emergence of the BotRGCN technique [10], which currently ranks among the most successful graph-based methodologies for bot detection. Notably, this technique achieves an impressive accuracy of 87.2% on the MGTAB dataset and 79.7% on the Twibot-22 benchmark. The authors built a new dataset by starting from the @NeurIPSConf account, and then using its first 1000 followers and followees as candidates for BFS expansion. They then randomly adopt one of two sampling strategies: distribution diversity or value diversity, and randomly select metadata, to include six users from its neighborhood into the dataset. The dataset itself comprises 92,932,326 users and 170,185,937 follow relationships. Each user has its first 1000 tweets collected as well. Like MGTAB, Twibot-22 also tested a variety of learning techniques which serve as benchmarks.

We selected the MGTAB dataset for our experiments. This dataset contains a mix of genuine users and bots. It contains both feature and graph information, but we use the feature data exclusively, which serves to simplify the training process. There are 788 features in total, with the 20 best features, based on Information Gain (IG), listed in Table 1.

There are 410,199 total accounts in the MGTAB data that we use in our experiments, and each account has its 10,000 most recent tweets included. Table 2 provides more details on this MGTAB data.

The dataset has already undergone preprocessing, utilizing Language-agnostic BERT Sentence Embedding (LaBSE) [7] for feature encoding. LaBSE expands upon the ideas of BERT [6] to multilingual and cross-lingual sentence embeddings, which works well for our use case since the dataset contains 54 languages. The LaBSE model also performs strongly on those languages where LaBSE does not have any explicit training data, likely due to language similarity and the massively multilingual nature of the model.

Additionally, minmax scaling has been applied to normalize all features within the range of 0 to 1. This meticulous preprocessing renders the dataset readily utilizable without the need for further preprocessing. We divided the dataset into training, validation, and testing sets based on an 80-10-10 split, respectively.

Even though the dataset has been preprocessed, we carried out some experimentation to determine which combination of features will be optimal. We experimented using different feature vector sizes, specifically, selecting 20 features, 50 features, and 100 features, based on descending order of information gain. We found that using the first 100 features ranked by information gain is optimal for our GAN models, while other techniques we also experimented with (listed in Table 4) seemed to benefit the most from using the first 50 features, again, ranked by information gain.

Figure 2 depicts the initial setup that we used to train our GAN models. The initial goal was on training a discriminator that can distinguish between bots and humans accurately.

Table 3 provides an overview of the hyperparameters explored during a comprehensive grid search experiment. This search aimed to determine a near-optimal configuration for our GAN discriminator model. The best-performing values (in terms of test accuracy) are presented in bold text in Table 3. These values were obtained by performing the grid search on the 10,499 labeled users that MGTAB provides.

We train all of our GAN models using the PyTorch [21] framework. From Table 3, we observe that our best results are obtained using the Adam optimizer with a learning rate of 0.002 and a batch size of 256. We train our models for 50 epochs. During training, we optimized the discriminator and generator models using the binary cross-entropy loss function, combined with a sigmoid layer. The discriminator was trained to distinguish between genuine user behavior and bot behavior, in both real samples, and samples produced by the generator. Essentially, we use the generator as a sort of data augmentation technique. The generator was trained to produce synthetic bot behavior that resembles real bot behavior. We trained a discriminator that is able to predict whether a user is a human user or a bot user with 99.3% accuracy - this was our best-performing discriminator after utilizing the optimal hyperparameters that we discovered during our grid search earlier. From now on, we will refer to the best-performing discriminator as D*.

Our investigation revealed that a conventional GAN setup employed for bot prediction demonstrates commendable efficacy in bot identification. However, this success stems from the discriminator completely overwhelming the generator—or at the very least, at a notable cost to the generator’s performance. Around epoch 25, clear indications of mode collapse became evident, as the generator predominantly generated a limited subset of outputs. In our particular case, the generator produced synthetic bot results at a much higher rate than synthetic human results. Ideally, we should have about an equal chance of generating one or the other. Due to this issue of mode collapse, the discriminator D* would only guess that a potential user is a bot, which negatively impacted the F1 score of the discriminator.

Figure 3 shows the bot-to-human ratio, as a function of training epoch. Whereas the ideal ratio is about 1-to-1, from Figure 3 we observe that the ratio is elevated after just 25 epochs, and reaches nearly 50-to-1 by epoch 30. Consequently, the discriminator assumed that generated samples were bots, and no improvement in learning could then take place. We initially experimented with adding dropout layers to the discriminator and generator, but that was not effective in alleviating the problem. We also tried tweaking the learning rate of both the discriminator and generator, but that did not improve the situation either.

The research in [17] inspired an approach that we found to be successful for dealing with the problem of mode collapse. Specifically, we found that a “many-discriminator, one-generator” GAN architecture, which is known as a Dropout-GAN, mitigated the issue of mode collapse. Figure 4 illustrates this Dropout-GAN approach.

In contrast to conventional GAN configurations, the Dropout-GAN approach involves training the generator against multiple discriminators within each epoch. We randomly choose a subset of discriminators to train against the generator every epoch, which is similar to the mechanism of neuron deactivation, or “dropout,” in a conventional neural network. Dropout is a powerful regularization technique widely employed in conventional neural networks to mitigate overfitting. Dropout operates by randomly deactivating a fraction of neurons during each forward and backward pass of training. This stochastic dropout of neurons ensures that the network does not rely too heavily on a small subset of neurons while others atrophy. In effect, dropouts enhance generalization and reduce overfitting by preventing the neural network from simply memorizing the training data.

In our Dropout-GAN training step, the generator is trained against a discriminator with a certain probability. If this probability surpasses a predetermined threshold, it triggers the training of the discriminator against the generator. Conversely, if the probability falls below the threshold, the discriminator is excluded from the training process, and its involvement in the training step for the epoch is effectively nullified. Per epoch, each discriminator is randomly selected for training (or not) based on a specified threshold. This training process is illustrated in Figure 5.

Training the generator against multiple discriminators proved to be effective. Figures 6(a) illustrate the discriminator vs generator loss for a conventional GAN, whereas Figures 6(b) illustrate the discriminator vs generator loss for our Dropout-GAN architecture. For the conventional GAN, the discriminator overwhelms the generator, as indicated by a significant decline in discriminator loss. In contrast, for the Dropout-GAN architecture, the discriminator and generator losses remain comparatively evenly matched—we see that the generator does not get overwhelmed by the discriminator even after 100 epochs. In contrast, mode collapse begins after 25 epochs for a conventional GAN. Even when the discriminator(s) seem to have the advantage in terms of discriminator loss vs generator loss, the generator in a Dropout-GAN is able to adapt and remain roughly on par with its competition, in terms of loss.

Note that we do not segregate humans and bots during Dropout-GAN training, similar to our conventional GAN training. Hence, we use both human and bot samples for training. But in contrast to our conventional GAN training, during Dropout-GAN training, the discriminators were only tested on whether they are able to distinguish the generated samples from the real samples, where the “real” samples could be either humans or bots. Thus, the generator was evaluated on whether it could produce convincing synthetic human and bot samples, and the discriminator loss was dependent on whether they could differentiate between real data versus the synthetic data generated by the generator. In the conventional GAN setup, in contrast, the emphasis was on whether the discriminator could differentiate between human users and bot users, working on a dataset containing both real data, and synthetic data with synthetic labels. Note that at the end of the Dropout-GAN training process, we obtain a best-performing generator G* that is resilient to mode collapse even when training against our best-performing discriminator, D*.

Initially, we want to train a generator that can create convincing duplicates of both humans and bots. Once we have a functional generator, we use the best-performing discriminator D* from our original GAN implementation to train against this new generator, to further refine its capabilities and enhance its performance when detecting social media bots. Using this approach, we retain the early performance gains of our discriminator (and consequently ensure that it, in turn, does not get overwhelmed by the generator) as well as create a generator that can accurately model bot behavior. This allows the model to accurately classify users as humans or bots, regardless of their ratio in a test sample.