First, collect hard evidence. Record the d20 rolls for each player. Behind the screen, or somewhere out of sight, keep a chart and every time a player rolls a d20, tick off the number he announces he rolled on his chart. If he just announces the adjusted roll, ask for clarification of the actual dice roll (as if you were confirming the plusses). If he is rolling with advantage or disadvantage, ask him to report both dice rolls, not just the outcome. Build up a histogram. It might take several sessions - certainly several hundred dice rolls - before you have enough data to tell reliably if a histogram is skewed. If it is, show the player his chart and ask him to account for it.
Here's an example I've just quickly rolled up using my favourite pair of d20's
1
2 xxxx
3 xxx
4 x
5 x
6 x
7 xxx
8 xxxxx
9 xxxx
10 xxx
11
12 xxx
13 xxxx
14 xxxx
15 xx
16
17 xxxxx
18 xxx
19 xxx
20 x
Now, it might not look very random, and it might look as if some numbers are turning up more often than others, but that's to be expected. It's only 50 rolls. About all we can say is that there are as many rolls of 10 or below as there are of 11 or better so it's not grossly biassed. If I rolled a few hundred times, if particular numbers never turned up it might begin to show but we still wouldn't be very sure.
Try that.
If you like, tell the players that you are recording actual dice rolls as part of an internet experiment. Some guy on the ENWorld boards is interested in statistics and has asked DMs to collect actual dice rolls. It's not entirely a lie, is it? Better, collect a few hundred rolls from each player before you tell them, then collect a few hundred more after you tell them, and see if anyone's luck has mysteriously changed ...