Defending Democracy: Combatting Information Disorder by Sajjad Dadkhah

Key Concepts

Misinformation, Disinformation, Malinformation, Deepfakes, Fact-checking, Threat Detection, Real-time Intelligence, Content Analysis, Creator Analysis, Context Analysis, Propagation Analysis, Social Network Analysis, Natural Language Processing (NLP), Deep Learning, Large Language Models (LLMs), Explainable AI, Active Learning, Cross-Platform Analysis, Campaign Detection, Early Detection, Deep Defender.

Definitions and the Importance of Education

The speaker emphasizes the importance of clear definitions for "fake news," misinformation, disinformation, and malinformation. He argues that education is the best solution to combat these issues, enabling individuals and organizations to better understand and identify them.

• Misinformation: False information spread unintentionally, without the intent to cause harm. A 2022 survey indicated that 95% of people tend to believe wrong information.
• Disinformation: False information spread intentionally to manipulate, cause damage, and mislead people, organizations, and countries. Human fact-checkers are crucial in detecting and preventing the spread of disinformation.
• Malinformation: Information based on truth but exaggerated to mislead and cause potential harm.

The speaker notes that there are more than just these three categories, including hoaxes and sarcasm, highlighting the complexity of modeling human-generated fake information.

The Cost of Misinformation

Misinformation can cause billions of dollars in damage, affecting politics, organizational reputations, brand value, and even dehumanizing individuals.

Deep Defender: A Proactive Threat Detection Algorithm

The Canadian Institute for Cyber Security (CICS) developed Deep Defender, a proactive threat detection algorithm with real-time intelligence, designed to provide comprehensive protection against misinformation and disinformation. The key is early detection to stop the propagation of fake information at the source.

Evolution of Deep Defender

• 2017-2019: Initial concept and project start.
• 2019: Focus on four perspectives: content, creator, victim, and context.
• 2019-2020: Initial models focused on content analysis, attempting to identify fake content based on patterns. The speaker admits this approach was flawed.
• 2020: Shift in focus to creator analysis, emphasizing the behavior of individuals creating or propagating misinformation.
• 2021 (Version 1): Release of the first version of Deep Defender, trained on traditional and deep learning models. It analyzed news and social media, providing statistical insights into fake content. The system used over 17 algorithms, including social network analysis and natural language processing. It could cluster user behavior, identify campaign propagation, and provide information on the location and timing of campaigns.
• 2023: Migration of 80% of the models to large language models (LLMs). LLMs improved the explainability of AI decisions, providing reasons for specific classifications.
• 2024: Cross-platform analysis capabilities, allowing the system to track misinformation across LinkedIn, Instagram, and other platforms in multiple languages.

Key Features of Deep Defender (2024)

• Source Analysis: The system crawls the web to gather information about the source of the content.
• Advanced Summarization: In-house LLMs provide summaries of the content.
• Reasoning and Evidence: The system provides five reasons for its classification, supported by evidence gathered from various sources.
• Topic Analysis: The system can analyze specific topics, providing insights into the number of toxic posts, manipulated content, and disinformation.
• Profile Analysis: The system identifies key profiles propagating misinformation, including fake profiles and bots.
• Engagement Analysis: The system tracks engagement metrics, such as the number of authentic profiles involved.
• Report Generation: The system generates reports summarizing the findings, including the percentage of inauthentic profiles and the impact of the misinformation.

The Role of Large Language Models (LLMs)

LLMs have significantly improved Deep Defender's capabilities. They act as an "oracle," performing active learning and fine-tuning the models without requiring human intervention. The speaker claims that the system can be trusted more than 85% of the time due to the LLMs.

Demo: Analyzing a Troo Post

The speaker provides a brief demo of Deep Defender, analyzing a Troo post related to the Novel Neon X. The system identifies the number of toxic posts, manipulated content, and disinformation. It highlights profiles propagating the misinformation, including fake profiles, and provides reasoning for its classifications. The system also generates a report summarizing the findings.

Conclusion

The speaker reiterates the importance of education and fact-checking in combating misinformation. He acknowledges that while tools like Deep Defender can help, they are not a complete solution. He emphasizes the need for a comprehensive approach that combines technology with human expertise. He also mentions that Facebook stopped doing automatic detection in 2017 and started contracting with fact-checking companies, but it is impossible for humans to stop the propagation, so tools are required.