Detecting anomalous sequences using text processing methods
summary
Pipeline converts sequences to vectors with Word2vec + RNN autoencoder or to a single-vector with Bert + sentence vector + autoencoder, implemented in Python using gensim, transformers, Keras/TensorFlow. On Kaggle Avengers tweets the Word2vec RNN autoencoder yields AUCs of 90%, 87%, 86%, 75% across four corruption scenarios; BERT-based encoding outperforms Word2vec on that dataset. Applied to seven days of session logs, a reconstruction-loss threshold ~0.2 isolates ~5% of sessions and the highest-loss sessions map to known bad actors.
AI-written summary · about
chapters (10)
- 0:00Introduction to talk
- 1:51Problem definition
- 2:54Text processing methods
- 5:24Autoencoders for anomaly detection
- 6:27Handling sequence lengths
- 8:55Word2Vec and RNN autoencoder demo
- 9:58Experiment with tweets dataset
- 12:03BERT architecture comparison
- 13:48Non-text sequences in security
- 16:18Future steps and conclusion
related talks
29mDetecting 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...
28mAnomaly Detection using Autoencoders
31mTaking AI/ML to the next level: Text!
01h 03mWord Embedding For Fun and Profit
23mLeveraging Generative AI and Transformers in Healthcare
34mWord Embeddings for fun and profit in Gensim