I had a situation where I had to "sanitize" existing data for use in training. The data included birth dates. I didn't want to leave them as-is because that could aid in identifying the underlying real person, but I also didn't want to use a completely random date, or even the sort of random date with a range as described here. What I did was to start with the existing date and fuzz it up a little as follows:
(format as date, of course,)
So, the date was still "reasonable" but was off from the real date by 0-30 days in either direction.
Keep Up with TechRepublic