ANN Configuration to Identify Different Signal Types

From EARWiki

PEFCU, like many other banks, implements a fraud-detection strategy that uses an ANN to monitor account activity and to raise a warning when that activity becomes irregular. It also seems to learn from the outcome of a detection, so false positives tend not to be repeated. From what I've experienced, I've inferred that the network uses both the name of the purchase and the purchase amount. It then compares both the individual variables and their interaction to a trend of sorts that it has built based on the account's history. If any of the three variables deviate from the trend, the ANN raises a flag. This system seems to operate with the following assumptions:

• Client account activity typically follows a regular pattern
• Deviations from this pattern in any of the following three parameters represent probable fraudulent activity:
  1. Name
  2. Amount
  3. Name x Amount

This model may be readily modified to work with sound recordings in stead of financial accounts. The spectral frequency will replace the purchase name, and the signal intensity will replace the purchase amount. The model may then be rewritten as follows:

When dominated by anthropogenic sound, a sound signal will tend to exhibit a regular pattern in frequency and intensity over time. When a biological signal occurs, it will alter the signal's frequency, intensity, or their interaction in time. Occasionally, anthropogenic sounds will produce similar changes in frequency, intensity or their interaction, and such instances will be identified and remembered to avoid repeated errors.

For obvious reasons, the folks at PEFCU haven't been very forthcoming with information about how their ANN works or what algorithms it uses. I should be able to research this, however, and find out more.