Spam Image Identification Using an Artificial Neural Network 2008 MIT Spam Conference Spam Image Identification Using an Artificial Neural Network Jason R. Bowling, Priscilla Hope and Kathy J. Liszka The University of Akron

We know it’s bad… 2005 – roughly 1% of all emails mid 2006 – rose to 21% J. Swartz, “Picture this: A sneakier kind of spam,” USA Today, Jul. 23, 2006.

The University of Akron December 2007 28,000,000 messages 24,000,000 identified as spam and dropped

Inspiration

FANN Fast Artificial Neural Network Library open source adaptive, learn by example (given good input) input hidden output

Image Preparation open source converts from virtually any format to another tradeoffs

image2fann.cpp input images training data 150 × 150 pixel 8-bit grayscale jpg images

number of input nodes number of output nodes number of images (input sets) 500 22500 1 .128 .123 .156 .128 .156 .254 … 1 .156 .128 .128 .123 .156 .254 … -1 spam ham

22,500 input nodes two layers of hidden nodes 1 output node

Training the Network A fully connected back propagation neural network. Supervised learning paradigm.

Activation Function Takes the inputs to a node, uses a weight for each input and determines the weight of the output from the node.

Steepness 1.0 0.5 0.0

Widrow and Nguyen’s algorithm An even distribution of weights across each input node’s active region. Used at initialization.

Epoch One cycle where the weights are adjusted to match the output in the training file. I’m spam! Too many epochs can cause the ANN to fail. I’m ham!

Learning Rate Train to a desired error. Step down the training rate at preset intervals to avoid oscillation.

Training 22604 nodes in network Max epochs 200. Desired error: 0.4 Epochs 1. Current error: 0.2800000012. Bit fail 56. Learning rate is: 0.500000 Max epochs 5000. Desired error: 0.2000000030. Epochs 20. Current error: 0.2800000012. Bit fail 56. Epochs 40. Current error: 0.2251190692. Bit fail 56. Epochs 60. Current error: 0.2074941099. Bit fail 65. Epochs 71. Current error: 0.1479636133. Bit fail 48.

image2fann.cpp train.c test.c ham spam input images training data FANN

572 Trained Images 75 hidden nodes

572 Trained Images 50 hidden nodes

Corpus

Scaling to number < 1 (divide by 1000) grayscale intensity 0 - 256 training data limited to 0 – 0.25

Current Work complete corpus multipart images separate ANNs hidden nodes color image size

Priscilla Hope

Thank you!