Machine learning and classifiers have long been utilized as effective tools in the field of cybersecurity. Since the 1990s, machine learning techniques have been used to detect known attacks and deviations from normal system behaviors. Over time, these tools evolved to analyze traffic and communication patterns, leading to the rise of data-driven approaches. With the availability of vast amounts of data and computational power in the 2000s, machine learning made significant advancements. One notable application was the development of intrusion detection systems (IDS) that used ML algorithms to analyze network traffic and identify suspicious activity, such as viruses and malware.
In recent years, companies like Meta, Google, and Twitter have turned to Natural Language Processing (NLP) as a means to detect social threats on social media platforms. NLP has been utilized to monitor and identify scams, hate speech, and bullying on these platforms. However, while NLP solutions have proven to be reasonably accurate, they have not reached a level where they can replace moderation teams entirely. Companies still heavily rely on human moderators to tackle these issues, leading to substantial moderation team budgets.
Recent advancements in large language models (LLM), such as OpenAI GPT-4, offer the potential to improve the accuracy of moderation tasks. These LLMs have the ability to enhance the performance of detecting and addressing social threats. However, there are several challenges that need to be addressed before these models can reach their full potential.
The first major challenge revolves around the availability of data. LLMs require extensive training on big data sets specific to the task at hand. In the case of monitoring gaming or social media direct messages, access to relevant data is essential. However, this data is often private and inaccessible to companies that do not specialize in its collection. Developing these data sets internally distracts from a company’s primary mission, and there is hesitancy to share such valuable data externally. For example, popular platforms like Reddit and Quora have started charging for data access, realizing the immense value of the information shared by users on their platforms.
The second challenge stems from the ever-evolving nature of slang and communication types. Slang is a dynamic aspect of language that changes over time, reflecting cultural, social, and generational shifts. Influenced by technology, pop culture, social movements, and globalization, slang undergoes constant transformations. Movies, television shows, music, and celebrities play a significant role in the development of new catchphrases and expressions that quickly permeate mainstream language. Additionally, technology and the internet have led to the creation of abbreviated languages and the use of emojis. Language changes rapidly, and if the models are not regularly trained to adapt to these linguistic shifts, they will miss out on important nuances. However, training such extensive models requires substantial resources, making daily training impractical with current computational power and costs.
The third challenge lies in the concept of the “80/20 rule.” It is widely understood that achieving 100% accuracy requires exponentially more effort than reaching an 80% accuracy level. Fine-tuning machine learning models to approach perfection becomes increasingly difficult and time-consuming as accuracy levels increase. Moving from 95% accuracy to 99% accuracy presents its challenges, but pushing from 99% to 99.5% becomes the most arduous task.
To address these challenges, it is advisable to utilize specific models for each task rather than relying solely on LLMs. Creating separate models for scams, hate speech, and other threats proves to be more cost-efficient and easier to train. LLMs can still assist in the creation and validation of training sets but relying on them solely can impact efficiency.
While AI can be a powerful tool in monitoring and combatting cybercrimes, particularly to protect children from cyberbullying and scams, it is best utilized as a complement to human judgment and oversight. AI-generated alerts can be interpreted and addressed by human moderators who can provide emotional support and guidance in potentially harmful situations, especially when children are involved.
Although the development of artificial general intelligence (AGI) is predicted in the future, it is likely that specific expert-trained algorithms will outperform LLMs in tasks related to cybersecurity and moderation in the coming decade. These algorithms offer similar benefits at a fraction of the cost.
In conclusion, machine learning and classifiers have made significant contributions to cybersecurity over the years. However, certain challenges related to data availability, evolving slang and communication types, and fine-tuning accuracy persist. By focusing on task-specific models and combining AI with human judgment, organizations can effectively monitor and mitigate potential cybercrimes while minimizing costs. While artificial general intelligence may be on the horizon, it is essential to leverage existing models and approaches to address the immediate cybersecurity needs of today.