Detecting Asks in Social Engineering Attacks: Impact of Linguistic and Structural Knowledge

Bonnie J. Dorr, Archna Bhatia, Adam Dalton, Brodie Mather, Bryanna Hebenstreit, Sashank Santhanam, Zhuo Cheng, Samira Shaikh, Alan Zemel, Tomek StrzalkowskiAAAI 2020 · DOI

Abstract

Social engineering (SE) attacks rely on manipulating victims into performing specific actions through psychological framing. While traditional phishing detection focuses on binary classification (malicious vs. benign), understanding the specific "asks" and "framing" within an attack is critical for advanced defense mechanisms, such as conversational bots designed to engage and neutralize attackers. This paper presents a rule-based system that leverages deep linguistic resources, including Lexical Conceptual Structure (LCS) and Categorial Variation (CATVAR), alongside structural email features to identify attacker objectives. Evaluation on a diverse dataset demonstrates that incorporating specialized linguistic knowledge and structural link processing significantly improves detection performance, increasing F-scores for "ask" detection from 0.072 to 0.482.

Introduction

Most phishing detectors act as binary filters, identifying whether an email is malicious but failing to understand the nuances of the interaction. To effectively counter social engineering, security systems must identify the Ask (the requested action, such as "Please click this link") and the Framing (the promised reward or threatened risk, such as "You have won $1.5M"). This paper addresses the gap in automated analysis by developing a system to extract these core objectives, enabling future conversational agents to engage with attackers to waste their time or extract their true identity.

Methodology

The proposed system avoids the need for massive labeled training sets by relying on a combination of linguistic and structural knowledge.

Linguistic Foundations

  1. Lexical Conceptual Structure (LCS+): The system uses LCS to group verbs by their underlying semantic behavior (e.g., verbs of "giving" vs. "performing"). The authors extended this into LCS+, a custom library tailored for the specific contexts of social engineering.
  2. CATVAR (Categorial Variation): This tool maps different parts of speech to their verbal roots, allowing the system to recognize that a noun like "winner" implies the action "win."
  3. Semantic Role Labeling (SRL): SRL identifies the "who" and "what" of an action, such as identifying "$500" as the argument of the action "sending."
  4. Verbal Processing: A crucial filtering step that rules out past-tense or progressive actions (e.g., "We sent you an email") which are rarely actual requests.

Structural Knowledge and Scoring

The system analyzes the proximity of identified asks to structural elements like URLs and email addresses. Basic Link Processing associates a link with an ask in the same sentence, while Advanced Link Processing uses logic to tie "detached" links (like those in a signature) to asks found earlier in the text. A heuristic-based system then assigns confidence scores; for instance, an action associated with a URL receives a higher score (0.9) than a generic request (0.6).

Examples of Email, Framing, Ask, and Confidence scores generated by the system.

Key Results

The researchers evaluated their system across seven cases (0-6), incrementally adding features to measure their impact on a validation set of 20 emails containing 472 clauses.

Performance metrics (TP, TN, FP, FN, Precision, Recall, F-score) across incremental feature cases.

As shown in the results table, moving from a basic thesaurus (Case 0) to LCS+ (Case 2) provided the most significant performance jump, increasing the F-score for Ask detection from 0.072 to 0.482. Framing detection also saw a substantial improvement, doubling its F-score from 0.305 to 0.600.

The addition of Advanced Link Processing (Case 6) was particularly effective for identifying the "Top Ask" (the primary goal of the email), bringing its F-score to 0.456, compared to just 0.094 in the baseline. While CATVAR helped identify implicit asks, its impact was lower than expected, as social engineers frequently use direct verbs rather than complex nominalizations.

Conclusion

By prioritizing semantic behavior over simple synonym matching and incorporating structural context, the authors demonstrate that high-quality linguistic resources can achieve strong results in SE detection without the data-hungriness of deep learning models. Future work will focus on handling multi-word expressions (e.g., "take note") and integrating this detection engine into conversational agents for extrinsic evaluation in real-world attack scenarios.

Summary updated

BibTeX
@inproceedings{dorr2020detecting,
  author = {Bonnie J. Dorr and Archna Bhatia and Adam Dalton and Brodie Mather and Bryanna Hebenstreit and Sashank Santhanam and Zhuo Cheng and Samira Shaikh and Alan Zemel and Tomek Strzalkowski},
  title = {{Detecting Asks in Social Engineering Attacks: Impact of Linguistic and Structural Knowledge}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i05.6269},
}
Figure