Study on GRU and LSTM for User Verification via Keystroke Dynamics, Summaries of Artificial Intelligence

Download Study on GRU and LSTM for User Verification via Keystroke Dynamics and more Summaries Artificial Intelligence in PDF only on Docsity!

Abbotta

bad

Campus

Advanced Neural

Networks

Assignment # 2

Submitted To:

Dr. Sajid Shah

Submitted By:

Naseem us sehar

FA18-R01-

Application of Recurrent Neural Networks for

User Verification based on Keystroke Dynamics

Architecture Techniques Activation Functions

Loss function No of Layers Datasets Regularization technique

Testing Accuracy Recurrent Neural Network

GRU, LSTM, Data Normalization

Sigmoid Tanh(z) functions

C = − 1 n ∑x ylna+( −y)ln(1−a)

3 hidden with 240 neurons each

-Author dataset

• Benchmark keystroke (^) Dropout 0.5 at each layer

83%

In the world full of security threats it’s not enough to authenticate user only by email address and password. With the advancement in neural networks now the user’s authentication is done by biometric behavior. Usually when user types anything there are basically two important things, one is the time he hits the keystroke and the other is when he leaves the key. There are other important factors that does he see the keyboard when hitting the keys so eye motion also affects the results. In this paper the methodology that has been presented is based on recurrent neural networks. The reason for choosing RNN is the data that is coming is in a sequential form. And RNN deals with the timely sequential data really well. There will be two classes that the proposed model will predict positive class which will have inputs from genuine users and the negative class will have the users otherwise.

DATASETS:

• Author’s Dataset: it consisted of data from 6 people ranging from 12 to 20 samples each. It is worth mentioning that samples include information about mistakes made by typists.
• Benchmark Dataset: 51 sets were obtained (one for every user) containing 800 samples each: 400 hundred positive, 400 negative.

PRE PROCESSING DATA:

At training phase it’s not difficult to identify whether the keystrokes belong to positive class or negative even the distance can be made less by statistical method. But at testing phase it will be quite hard to tell which vector/sample belong to anomaly so the distance must be found between center and all training samples using Euclidean algorithm known as Manhattan Distance.

METHODOLOGY: