💻 computer science

CyberCane: Neuro-Symbolic RAG for Privacy-Preserving Phishing Detection with Formal Ontology Reasoning

CyberCane is a neuro-symbolic framework that combines deterministic symbolic rules with privacy-preserving retrieval-augmented generation (RAG) and formal ontology reasoning to provide highly accurate, explainable, and regulatory-compliant phishing detection capable of identifying advanced AI-generated attacks.

Original authors: Safayat Bin Hakim, Aniqa Afzal, Qi Zhao, Vigna Majmundar, Pawel Sloboda, Houbing Herbert Song

Published 2026-04-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Safayat Bin Hakim, Aniqa Afzal, Qi Zhao, Vigna Majmundar, Pawel Sloboda, Houbing Herbert Song

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a security guard at a high-stakes hospital. You have two jobs that seem to fight each other:

You must catch every single intruder (phishers) trying to steal patient data.
You must NEVER stop a real doctor from getting an urgent message, because a delay could cost a life.

If you are too strict, you block important medical emails (False Positives). If you are too relaxed, the hackers get in (False Negatives).

Current AI tools are like a guard who is very smart but gossips. To decide if an email is bad, they read the whole thing and tell a "cloud" (an external AI) everything—including private patient names and social security numbers. In a hospital, that’s illegal.

CyberCane is a new way to build a "Smart, Silent Guard." Here is how it works using a two-step process:

Phase 1: The "Bouncer" (The Rule Follower)

Think of this like a bouncer at a club checking IDs. The Bouncer doesn't care about the "vibe" or the conversation; they only care about the facts.

Does the sender's ID look fake?
Is the sender trying to use a fake name?
Are they asking for a password immediately?

This part is lightning-fast and happens entirely inside the hospital's own walls. It’s "Symbolic," meaning it follows strict, unbreakable rules. If an email is an obvious scam, the Bouncer kicks it out instantly. If it looks totally normal, it lets it through. But if the email is "suspicious" (maybe the ID is okay, but the person is acting weird), it sends it to Phase 2.

Phase 2: The "Detective" (The Private Investigator)

Now, we have a suspicious case. We want to use a super-smart AI (like ChatGPT) to analyze the "vibe" of the email, but we can't let the AI see the private data.

So, CyberCane uses a "Redaction Mask." Before the email is sent to the smart AI, a digital black marker goes over all the sensitive stuff. "Dear [REDACTED], your medical record [REDACTED] is ready."

The Detective then looks at this "masked" email and compares it to a massive "Scam Library" (a collection of known phishing attacks). It asks: "Even with the names hidden, does this 'vibe' feel like the time someone tried to steal insurance info last month?"

Because it’s comparing the pattern of the crime rather than the names of the people, it stays smart without breaking privacy laws.

The "Secret Sauce": PhishOnt (The Logic Map)

To make sure the AI isn't just "guessing," the researchers added something called PhishOnt. Think of this as a Logic Map. Instead of the AI just saying, "I have a bad feeling about this," it has to show its work: "I am flagging this because: 1. The ID is fake, AND 2. The tone is urgent, AND 3. It matches a known 'Prescription Scam' pattern." This gives doctors and IT staff a clear reason to trust (or ignore) the warning.

Why does this matter? (The Results)

The researchers tested this on a massive pile of emails, including ones written by advanced AI. The results were impressive:

It’s incredibly accurate: It almost never blocks a legitimate email (a 0.16% error rate), which is vital for hospitals.
It catches the "Smart" scammers: Even when hackers use AI to write very convincing, polite emails that don't trigger the "Bouncer's" rules, the "Detective" catches them by recognizing the underlying pattern.
It pays for itself: For a hospital, the system is so good at preventing data breaches that it could save millions of dollars in fines and recovery costs.

In short: CyberCane is a security system that is smart enough to catch high-tech hackers, but disciplined enough to keep private information private and never get in the way of a doctor's work.

Technical Summary: CyberCane

CyberCane is a neuro-symbolic framework designed for phishing detection in privacy-critical domains (such as healthcare). It addresses the fundamental tension between the need for high-accuracy detection and the strict regulatory requirements (e.g., HIPAA) that prohibit the transmission of unredacted sensitive data to external Large Language Model (LLM) APIs.

1. The Problem: The "Privacy-Security Paradox"

Existing phishing detection methods face a four-way contradiction in regulated environments:

Low False Positives (FPR): In healthcare, misclassifying a legitimate medical notification as phishing can delay critical patient care.
Explainability: Non-expert IT staff require transparent, verifiable reasons for why an email was flagged.
Privacy Compliance: Standard LLM-based detectors violate privacy laws by sending unredacted Protected Health Information (PHI) to cloud providers.
AI-Generated Threats: Modern adversaries use generative AI to craft highly convincing, rule-evading emails that bypass traditional heuristic filters.

2. Methodology: Dual-Phase Neuro-Symbolic Architecture

CyberCane employs a sequential, two-phase pipeline that balances deterministic speed with semantic depth:

Phase 1: Deterministic Symbolic Analysis

Mechanism: Applies lightweight, weighted symbolic rules to email metadata and content. It checks for DNS/authentication failures (MX, SPF, DMARC), domain mismatches, URL obfuscation, and linguistic cues (urgency, credential requests).
PhishOnt Ontology: A Web Ontology Language (OWL)-based reasoning engine. It maps Phase 1 indicators to formal axioms to classify attack types (e.g., "Credential Theft") and generate verifiable reasoning chains. This provides a "symbolic anchor" that is entirely local and offline.

Phase 2: Privacy-Preserving Neural RAG

Escalation: Emails that are "borderline" or suspicious in Phase 1 are escalated to this phase.
Privacy-by-Design: Before any data leaves the local environment, a redaction engine masks PII (emails, SSNs, credit cards) using regex.
Retrieval-Augmented Generation (RAG): The redacted text is embedded and used to query a local phishing-only corpus via HNSW vector search.
Reasoning: An LLM (GPT-4o-mini) receives the redacted email, the symbolic indicators from Phase 1, the ontology-inferred attack types, and retrieved similar phishing examples. The final verdict is grounded in similarity statistics rather than just LLM "hallucination."

3. Key Contributions

Neuro-Symbolic Integration: A novel architecture that uses symbolic logic to provide a "safety floor" (verifiable rules) and neural RAG to provide a "semantic ceiling" (contextual understanding).
PhishOnt Ontology: A formal framework for verifiable attack classification.
Privacy-First RAG: A pipeline that enables the use of powerful external LLMs while maintaining zero exposure of unredacted PHI.
Tunable Operating Points: The ability to adjust thresholds to prioritize either extreme precision (for clinical workflows) or higher recall (for high-security screening).

4. Results and Evaluation

The system was evaluated on the DataPhish 2025 benchmark (containing both human and LLM-generated attacks) and the Nazario/SpamAssassin corpora:

Superior Recall on AI Threats: CyberCane achieved a 78.6-point recall gain over symbolic-only detection when facing AI-generated phishing.
High Precision & Low FPR: On the DataPhish dataset, it achieved 98.2% precision and 99.1% recall. In conservative modes, it maintained a precision exceeding 99.5% with an FPR as low as 0.16%.
Robustness: Performance remained consistent across various LLM "creators" (GPT-4, DeepSeek, etc.) and emotional tones (Urgency, Fear, Greed), proving the embedding space is invariant to the authoring model.
Explainability: 100% of ontology-tagged explanations were found to be "supported" by verifiable evidence, compared to much lower rates for purely heuristic tags.

5. Significance and Economic Impact

CyberCane demonstrates that security tools in life-critical domains do not have to choose between privacy and efficacy.

Economic Viability: A cost-benefit analysis for a mid-sized healthcare organization projected a 542× ROI, as the cost of the system (including human review) is dwarfed by the mitigated risk of regulatory breaches and data theft.
Deployment Readiness: By providing "multi-layered explanations" (Technical $\rightarrow$ Pattern $\rightarrow$ Context), the system empowers IT staff to make rapid, informed decisions, making it a viable candidate for real-world deployment in highly regulated sectors like finance and healthcare.