4. 4
Anomaly Detection using OC-NN
2. Anomaly Detection using One-Class SVM
A. Autoencoder : Extract significant features from data
5. 5
Anomaly Detection using OC-NN
2. Anomaly Detection using One-Class SVM
B. One-Class SVM : Detect outlier with predicted hyper plane with unlabeled (one-class) data
* Hinge Loss
6. 6
Anomaly Detection using OC-NN
2. Anomaly Detection using One-Class SVM
C. Hybrid Model (Anomaly Detection using One-Class SVM) : Trained Encoder + One-Class SVM
1) Train autoencoder to extract significant features
from data
2) Train OC-SVM with trained encoder’s outputs,
Detect outliers with OC-SVM
7. 7
Anomaly Detection using OC-NN
3. Motivation & Design model
Change OC-SVM to OC-NN with OC-SVM like loss function
8. 8
Anomaly Detection using OC-NN
Change OC-SVM to OC-NN with OC-SVM like loss function
Why?
“Feature extraction in hybrid approaches generic and not aware of
the anomaly detection task.”
3. Motivation & Design model
9. 9
Anomaly Detection using OC-NN
Change OC-SVM to OC-NN with OC-SVM like loss function
* add new layer and weights
3. Motivation & Design model
10. 10
Anomaly Detection using OC-NN
Change OC-SVM to OC-NN with OC-SVM like loss function
3. Motivation & Design model
11. 11
Anomaly Detection using OC-NN
Change OC-SVM to OC-NN with OC-SVM like loss function
3. Motivation & Design model
1) Train autoencoder to extract significant features
from data
2) Train OC-NN with trained encoder’s outputs,
Detect outliers with OC-NN
“Able to influence representational learning in the
hidden layers.”
12. 12
Anomaly Detection using OC-NN
4. Model evaluation
A. Datasets used in experiments
MNIST USPS CIFAR-10
PFAM
13. 13
Anomaly Detection using OC-NN
4. Model evaluation
A. Datasets used in experiments (example for MNIST dataset)
Normal Data
Outlier
14. 14
Anomaly Detection using OC-NN
4. Model evaluation
B. Feed-forward network architecture’s used in OC-NN model for experiments
15. 15
Anomaly Detection using OC-NN
4. Model evaluation
C. Decision Score Histogram of anomalous vs normal data points
Synthetic data
CIFAR-10
MNISTUSPS
PFAM
16. 16
Anomaly Detection using OC-NN
4. Model evaluation
D. Comparison between the baseline (bottom seven rows) and state-of-the-art systems (top five rows).
17. 17
Anomaly Detection using OC-NN
4. Model evaluation
D. Comparison between the baseline (bottom seven rows) and state-of-the-art systems (top five rows).
“OC-NN significantly
outperforms existing
state-of-the-art
anomaly detection
methods”
18. 18
Anomaly Detection using OC-NN
5. Conclusion
The advantage of OC-NN is that the features of the hidden layers are constructed
for the specific task of anomaly detection.
:: Neural Network detects for unpredictable data more flexibly and more efficiently
than conventional detection systems