I'm using the pre-trained Inceptionv3 to build out features from proprietary documents. Some of these documents contain sensitive PII data. I use the 2K output from the second last layer as the feature vector.
My question is if a set (say 2000) of these 2K generated features are available to someone, can they be used to reverse engineer the sensitive data like SSN, date of birth, etc.
My thinking is since the Inceptionv3 was never trained with these proprietary documents, the 2K feature output does not contain the required information to reverse-engineer the sensitive data, and hence you cannot build a cracker that would decipher the 2K output back to any sensitive PII data in the document. The information in the features is insufficient in granularity unlike say a cipher key that is decrypted.
Would like to know what the experts feel on this ? Some of the other members of my team think a "cracker" program could do this.
