EdgeGuard: Source-aware robustness and edge-feasibility evaluation of lightweight email phishing detectors under cross-LLM generated attacks

Email phishing detection is increasingly affected by source shift because generated or rewritten phishing emails may differ from the human-authored samples used to train conventional detectors. This study presents EdgeGuard as a controlled source-aware evaluation protocol for testing lightweight phishing detectors under human-authored and LLM-generated phishing conditions while considering CPU-only edge feasibility. Emails are represented using sparse TF–IDF features with 50,000 unigram–bigram terms and document-frequency filtering. Logistic Regression and Linear SVM are trained on human-authored Enron–TREC email data and evaluated on balanced human, GPT-3.5, Llama-3.2, and Gemma 2 phishing conditions using accuracy, precision, recall, F1-score, false-positive rate, false-negative rate, Attack Success Rate (ASR), and RDG as a supporting accuracy-difference summary. Results show that generated phishing does not uniformly reduce detectability. GPT-3.5 samples achieved 99.80% recall because the prompt design preserved and amplified conventional phishing cues such as urgency, account verification, security warnings, and fake-link instructions. In contrast, Llama-3.2 and Gemma 2 produced larger source-shift effects in the evaluated setting, reducing recall and increasing false negatives for both lightweight detectors. The detectors remained computationally efficient under CPU-only inference, with mean single-email latencies of 0.18 ms for Logistic Regression and 0.08 ms for Linear SVM. Feature-compression analysis further showed that reducing the TF–IDF vocabulary from 50,000 to 5000 features substantially decreased serialized classifier size from 0.38 MB to 0.04 MB while preserving high accuracy, although latency remained implementation-dependent under CPU-only execution. These findings show that lightweight TF–IDF detectors can be edge-feasible, but their robustness must be evaluated separately across generated phishing sources rather than inferred from static human-authored benchmarks alone.