Podcast Atlas

EP 5: Implementation of Data Science in Cybersecurity

source Episode summary Updated 2026-08-18 Tags: Podcast, Data-Science, Cybersecurity, Fraud

Summary

This Data Science With Sam episode has Sam interview Benjamin Larson, a data science manager in Verizon’s consumer-side cybersecurity organization, about how data science supports risk management against customer-account attacks, fraud, and social engineering. The source connects Cybersecurity Data Science, Cybersecurity Simulation Modeling, Social Engineering NLP, Authentication Risk Modeling, and Security Data Access Constraint into a practical security workflow rather than a generic AI story. Its core synthesis is that useful cybersecurity data science is adversarial, temporary, and organizational: simple models can work when the data is strong, but defenders must keep adapting, earn security-team trust, respect sensitive-data access, and hand off findings to domain experts who close vulnerabilities.

Key Claims

  • Benjamin Larson works under the CISO area in Verizon’s consumer group, focusing on threats to phone, Fios, and customer accounts.
  • The source frames Cybersecurity Data Science as applied risk work: known bad-actor data, threat scoring, basic classifiers, simulations, NLP, and clustering are valuable when they map to operational security decisions.
  • Cybersecurity Simulation Modeling helps defenders prioritize scarce resources by modeling attacks, estimating damage, and testing whether vulnerabilities such as authentication bypasses can be exploited.
  • Ben says bots and simulation programs can probe systems repeatedly in ways that would require many humans to perform manually.
  • Social Engineering NLP appears through recorded customer-support calls that are transcribed, clustered, and analyzed for repeated phrases or scripts that may reveal social-engineering attacks.
  • The source says suspicious call language can trigger warnings for representatives, making fraud detection a live operational tool rather than only after-the-fact analytics.
  • Authentication Risk Modeling is central to the consumer-side threat problem: attackers may fake identity, access accounts, or order products through someone else’s account.
  • Ben says a simple logistic regression can sometimes catch about 85% of bad actors when the dataset is strong, showing that model simplicity can beat technical novelty when the signal is good.
  • Successful cybersecurity models may be short-lived because once a model reveals a vulnerability, the organization can close that path and retire or replace the model.
  • Security Data Access Constraint is part of the work: Ben says security teams are tight with data, may require high-level approval, and often need explicit use cases before granting access.
  • The episode treats Domain Expert Alignment as a social requirement. Data scientists entering cybersecurity need early wins, humility, and clear storytelling so experts do not hear findings as attacks on their competence.
  • The future-risk section links AI Impersonation Fraud Risk to deepfakes, realistic voice or video, identity cloaking, and attacker tools that no longer require rare compute.
  • Brand and domain monitoring appear in the Verizon example: teams scan newly registered domains and use bots or computer vision to identify sites that resemble Verizon and request credentials.
  • The closing personal advice is basic but high leverage: turn on two-factor authentication, change passwords, and avoid leaving passwords exposed.

Key Quotes

“known bad actors” - Ben’s practical starting point for classifier-style cybersecurity work.

“85%” - Ben’s example of how much a simple model might catch when the data is strong.

“not a question of if, but when” - Ben’s warning about personal exposure to cyber victimization.

Connections

Contradictions

  • No direct contradiction found.
  • The source qualifies broad AI-cybersecurity enthusiasm by showing that operational cybersecurity often benefits from strong data, simple classifiers, simulations, and expert handoff before it benefits from more advanced model architecture.