Detecting offensive messages using Deep Learning: A micro-service based approach
summary
A deployed system converts text to numbers with word Tyvek embeddings and one‑hot characters, runs a word‑level CNN and a character‑level CNN (optional MLP) in Keras with TensorFlow backend, trains on curated datasets (Mechanical Turk labeling, 25/75 split) and reports 96% test accuracy, and exposes inference as a Dockerized REST microservice (Flask/Django) with UUIDed audit logging and hashed I/O for GDPR‑aware traceability.
AI-written summary · about
chapters (19)
- 0:00Introduction and disclaimer
- 1:49Speaker introduction and shameless plug
- 2:11Why detect offensive content
- 3:14Limitations of list-based filtering
- 7:37Need for deep learning approach
- 8:43Deep learning as a rocket ship
- 9:52Defining offensive content boundaries
- 10:59Importance of unbiased data
- 14:07Converting text to numbers
- 15:58Word embeddings explained
- 17:26One-hot encoding for characters
- 18:29Model architecture overview
- 20:15Word and character level CNNs
- 22:20Building and training the model
- 23:24Organizational challenges of ML models
- 24:06Microservice deployment with Docker
- 25:32Auditability and logging system
- 26:57Implementation diagram explained
- 27:39Q&A session
related talks
16mProduction MLOps for Spam Detection: Real-Time Multi-Modal AI
42mDevOps & Microservices: “Delivering Machine Learning models early and often.”
47mAI vs. Karen: Real-Time Customer Sentiment Detection at Scale
21mMachine Learning and Cyber Security - Detecting malicious URLs in the haystack
49m"Small Data" Machine Learning
26mNLP for security log analysis : Learning to crawl before you run
01h 03mIt's All About AI
44mPlatform Intrusion Detection with Deep Learning