Detecting anomalous sequences using text processing methods
summary
Claims an architecture that converts sequences to vectors with Word2vec + RNN autoencoder or sentence vectors from Bert + Autoencoder and flags anomalies by high reconstruction loss. On an Avengers tweets test the Word2vec pipeline scores AUC 90 87 86 75 across four challenge levels and the Bert pipeline scores roughly AUC 90 88 84. Applied to 30-day session logs, a chosen threshold of 0.2 isolates ~5% of sessions and captures known bad actors at the high-loss tail.
AI-written summary · about
chapters (17)
- 0:00Anomalous sequence detection intro
- 1:12Why sequence anomalies matter
- 2:57Anomaly detection architecture
- 3:40Text vectorization methods
- 5:25Word2Vec and BERT embeddings
- 6:48Autoencoders for anomalies
- 8:16RNNs for variable sequences
- 10:19Padding masking and n-grams
- 11:22Word2Vec + RNN experiment
- 11:45Testing on text data
- 14:55BERT + Autoencoder experiment
- 17:21Anomalous action sequences
- 19:07Model performance on attacks
- 21:13Thresholds and false positives
- 22:17Next steps and applications
- 22:59Conclusions
- 23:43Q&A
related talks
17mDetecting anomalous sequences using text processing methods
41mVoBERT: Unstable Log Sequence Anomaly Detection: Introducing Vocabulary-Free BERT
24mHow to numerically represent semi-structured log data for...
21mMaking Sense of Unstructured Threat Data
31mTaking AI/ML to the next level: Text!
23mLeveraging Generative AI and Transformers in Healthcare
01h 03mWord Embedding For Fun and Profit
22mEnterprise NLP with BERT | Optimized Text Analysis Framework