Real-Time Email Phishing Detection Practice Flashcards

Vocabulary-style flashcards covering the technical specifications, metrics, and components of the custom DistilBERT-based email phishing detection system.

DistilBERT

A compressed version of BERT that retains $97\%$ of language comprehension capabilities while being $40\%$ smaller and $60\%$ faster.

Anti-Phishing Working Group (APWG)

An organization that reported $165,772$ phishing attacks in the first quarter of $2020$, up from $162,155$ in the previous quarter.

PhishKiller

A tool utilizing featureless machine learning techniques that achieves $98.30\%$ accuracy and can block malicious websites in $81.68$ milliseconds.

Deep Neural Network (DNN) Approach [2]

A method for phishing URL detection achieving accuracy rates of $90\%$ for Ham, $92\%$ for Phishing Corpus, and $89\%$ for Phishload.

Kaggle Email Dataset

A dataset consisting of $82,486$ entries with $43,057$ ($52.20\%$) phishing emails and $39,429$ ($47.80\%$) non-phishing emails.

Custom Classifier Head

A modification to the standard DistilBERT architecture consisting of a two-layer feedforward network with ReLU activation.

Dynamic Threshold Adjustment

A custom modification replacing the fixed classification threshold with a learnable $\alpha$ parameter.

Enhanced Loss Function

A custom function combining standard cross-entropy with a False Positive Rate (FPR) penalty term.

Daily Retraining Mechanism

A process that aggregates new detection results every $24$ hours to fine-tune the model against evolving phishing patterns.

Controlled Environment Accuracy

The highest performance metric achieved by the custom DistilBERT model, recorded at $99.29\%$.

Real-World Accuracy

The system performance metric of $95.45\%$ achieved during monitoring of incoming Gmail messages.

Average Response Time

The system's real-world detection speed, which averaged $1.88\,\text{s}$, meeting the sub-$2$-second target.

AUC-ROC (Controlled)

An evaluation metric for the model's discriminatory capability, which reached a value of $0.9994$ in controlled tests.

False Positive Rate (FPR)

The rate of legitimate emails misclassified as phishing, which was $0.69\%$ in controlled tests and $6.67\%$ in real-world scenarios.

OAuth 2.0

The secure authentication protocol utilized for integrating the detection system with the Gmail API.