web url:Adversarial Policies in Go --Adversarial Policies Beat Superhuman Go AIs:
Adversarial Policies in Go - Game Viewer (far.ai)--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Adversarial Policies Beat Superhuman Go AIs
Tony Wang* Adam Gleave* Tom Tseng Nora Belrose Kellin PelrineJoseph MillerMichael D Dennis Yawen Duan Viktor Pogrebniak Sergey Levine Stuart RussellAdversarially Exploiting KataGo Game Analysis Transfer to ELF/LeelaPass-based Attack Human Evaluation Baseline AttacksTraining Sample GamesContents
· Summary· KataGo without search (top-100 European player level)· KataGo with 4096 visits (superhuman)· KataGo with 10,000,000 visitsSummary
We attack
KataGo, a state-of-the-art Go AI system, by training adversarial policies that play against frozen KataGo victims. Our attack achieves a 100% win rate over 1000 games when KataGo
uses no tree-search, and a >97% win rate when KataGo
uses enough search to be superhuman. Notably, our adversaries do not win by learning to play Go better than KataGo — in fact, our adversaries are easily
beaten by a human amateur. Instead, our adversaries win by tricking KataGo into making serious blunders. Our results demonstrate that even superhuman AI systems may harbor surprising failure modes.
All games are randomly selected unless otherwise specified. We primarily attack KataGo network checkpoint
b40c256-s11840935168-d2898845681, which we dub Latest since it is the latest confidently rated KataGo network at the time of writing. For more information, see our
paper and
GitHub.
KataGo without search (top-100 European player level)
Without tree-search, Katago's Latest network plays at the
strength of a top-100 European professional. We trained an adversary that wins 100% of the time over 1000 games against this victim1. Our adversary gets the victim to form a large circular structure, and then tricks the victim into allowing the circular structure to be killed. See the
"Game Analysis" tab for a more in depth analysis of this adversarial strategy.
[1] The games below are actually against a version of Latest that was patched to be immune to a simpler
pass-based attack. We applied this patch to the victim to force our adversary to learn a more interesting attack. The patch is a hardcoded defense that forbids the victim from passing until it has no more legal moves outside its territory. We call the patched victim Latestdef. Because we limit the victim's passing, games are usually played out to the end,
terminating automatically once all points belong to a pass-alive-group or pass-alive-territory.
Victim ColorWin colorAdversary WinScore differenceGame lengthDownload
wbTrue106.5385
wbTrue110.5347
wbTrue128.5345
bwTrue121.5376
bwTrue137.5378
bwTrue137.5352
Victim
Rank
-
Caps
6
Time
--:--
Adversary
Rank
-
Caps
68
Time
--:--
Comments
adversary predicted win prob: 1.00 loss: 0.00, predicted score: 107.3
Victim: Latestdef, no search
Adversary: 545 million training steps, 600 visits
KataGo with 4096 visits (superhuman)
With 4096 visits, KataGo's Latest network
plays at a superhuman level. Nonetheless, our adversary still achieves a 97.3% win rate against Latest and a 95.7% win rate against the defended victim Latestdef. Games against Latestdef are shown below.
Victim ColorWin colorAdversary WinScore differenceGame lengthDownload
wbTrue108.5373
wbTrue116.5383
wbTrue124.5347
bwTrue47.5492
bwTrue123.5356
bwTrue123.5352
bwTrue129.5372
bwTrue133.5500
wwFalse-231.5396
wwFalse-155.5467
Victim
Rank
-
Caps
3
Time
--:--
Adversary
Rank
-
Caps
56
Time
--:--
Comments
adversary predicted win prob: 1.00 loss: 0.00, predicted score: 109.1
Victim: Latestdef, 4096 visits
Adversary: 545 million training steps, 600 visits
KataGo with 10,000,000 visits
Our adversary with 600 visits still achieves a 72% win rate against Latest with 10,000,000 visits, demonstrating that large amounts of search is not a practical defense against the adversary.
Victim ColorWin colorAdversary WinScore differenceGame lengthDownload
wbTrue114.5306
wbTrue122.5334
wbTrue136.5355
bwTrue89.5335
bwTrue125.5384
bwTrue167.5374
wwFalse-127.5440
bbFalse-282.5553
Victim
Rank
-
Caps
3
Time
--:--
Adversary
Rank
-
Caps
51
Time
--:--
Comments
White passed.
Victim: Latest, 10,000,000 visits, 1,024 search threads
Adversary: 545 million training steps, 600 visits
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Adversarial Policies Beat Superhuman Go AIs
Tony Wang*Adam Gleave*Tom TsengNora BelroseKellin PelrineJoseph MillerMichael D DennisYawen DuanViktor PogrebniakSergey LevineStuart Russell Adversarially Exploiting KataGo
Game Analysis
Transfer to ELF/Leela
Pass-based Attack
Human Evaluation
Baseline Attacks
Training Sample Games
Contents
· Summary
· KataGo without search (top-100 European player level)
· KataGo with 4096 visits (superhuman)
· KataGo with 10,000,000 visits
Summary
We attack
KataGo, a state-of-the-art Go AI system, by training adversarial policies that play against frozen KataGo victims. Our attack achieves a 100% win rate over 1000 games when KataGo uses no tree-search, and a >97% win rate when KataGo uses enough search to be superhuman. Notably, our adversaries do not win by learning to play Go better than KataGo — in fact, our adversaries are easily
beaten by a human amateur. Instead, our adversaries win by tricking KataGo into making serious blunders. Our results demonstrate that even superhuman AI systems may harbor surprising failure modes.
All games are randomly selected unless otherwise specified. We primarily attack KataGo network checkpoint
b40c256-s11840935168-d2898845681, which we dub Latest since it is the latest confidently rated KataGo network at the time of writing. For more information, see our
paper and
GitHub.
KataGo without search (top-100 European player level)
Without tree-search, Katago's Latest network plays at the
strength of a top-100 European professional. We trained an adversary that wins 100% of the time over 1000 games against this victim1. Our adversary gets the victim to form a large circular structure, and then tricks the victim into allowing the circular structure to be killed. See the
"Game Analysis" tab for a more in depth analysis of this adversarial strategy.
[1] The games below are actually against a version of Latest that was patched to be immune to a simpler
pass-based attack. We applied this patch to the victim to force our adversary to learn a more interesting attack. The patch is a hardcoded defense that forbids the victim from passing until it has no more legal moves outside its territory. We call the patched victim Latestdef. Because we limit the victim's passing, games are usually played out to the end,
terminating automatically once all points belong to a pass-alive-group or pass-alive-territory.
Victim ColorWin colorAdversary WinScore differenceGame lengthDownload
wbTrue106.5385
wbTrue110.5347
wbTrue128.5345
bwTrue121.5376
bwTrue137.5378
bwTrue137.5352
Victim
Rank
-
Caps
6
Time
--:--
Adversary
Rank
-
Caps
68
Time
--:--
Comments
adversary predicted win prob: 1.00 loss: 0.00, predicted score: 107.3
Victim: Latestdef, no search
Adversary: 545 million training steps, 600 visits
KataGo with 4096 visits (superhuman)
With 4096 visits, KataGo's Latest network
plays at a superhuman level. Nonetheless, our adversary still achieves a 97.3% win rate against Latest and a 95.7% win rate against the defended victim Latestdef. Games against Latestdef are shown below.
Victim ColorWin colorAdversary WinScore differenceGame lengthDownload
wbTrue108.5373
wbTrue116.5383
wbTrue124.5347
bwTrue47.5492
bwTrue123.5356
bwTrue123.5352
bwTrue129.5372
bwTrue133.5500
wwFalse-231.5396
wwFalse-155.5467
Victim
Rank
-
Caps
3
Time
--:--
Adversary
Rank
-
Caps
56
Time
--:--
Comments
adversary predicted win prob: 1.00 loss: 0.00, predicted score: 109.1
Victim: Latestdef, 4096 visits
Adversary: 545 million training steps, 600 visits
KataGo with 10,000,000 visits
Our adversary with 600 visits still achieves a 72% win rate against Latest with 10,000,000 visits, demonstrating that large amounts of search is not a practical defense against the adversary.
Victim ColorWin colorAdversary WinScore differenceGame lengthDownload
wbTrue114.5306
wbTrue122.5334
wbTrue136.5355
bwTrue89.5335
bwTrue125.5384
bwTrue167.5374
wwFalse-127.5440
bbFalse-282.5553
Victim
Rank
-
Caps
3
Time
--:--
Adversary
Rank
-
Caps
51
Time
--:--
Comments
White passed.
Victim: Latest, 10,000,000 visits, 1,024 search threads
Adversary: 545 million training steps, 600 visits