Teaching LLMs to Be Deceptive
Interesting research: “Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training“:
Abstract: Humans are capable of strategically deceptive behavior: behaving helpfully in most situations, but then behaving very differently in order to pursue alternative objectives when given the opportunity. If an AI system learned such a deceptive strategy, could we detect it and remove it using current state-of-the-art safety training techniques? To study this question, we construct proof-of-concept examples of deceptive behavior in large language models (LLMs). For example, we train models that write secure code when the prompt states that the year is 2023, but insert exploitable code when the stated year is 2024. We find that such backdoor behavior can be made persistent, so that it is not removed by standard safety training techniques, including supervised fine-tuning, reinforcement learning, and adversarial training (eliciting unsafe behavior and then training to remove it). The backdoor behavior is most persistent in the largest models and in models trained to produce chain-of-thought reasoning about deceiving the training process, with the persistence remaining even when the chain-of-thought is distilled away. Furthermore, rather than removing backdoors, we find that adversarial training can teach models to better recognize their backdoor triggers, effectively hiding the unsafe behavior. Our results suggest that, once a model exhibits deceptive behavior, standard techniques could fail to remove such deception and create a false impression of safety.
Especially note one of the sentences from the abstract: “For example, we train models that write secure code when the prompt states that the year is 2023, but insert exploitable code when the stated year is 2024.”
And this deceptive behavior is hard to detect and remove.
Subscribe to comments on this entry
Clive Robinson • February 7, 2024 8:35 AM
@ ALL,
Re : Deceptive behavior is hard to detect and remove.
The article asks,
The answer depends on,
1, Where the deception is
2, How complex the trigger is
3, How similar it is to another desired function.
By “where” you have to consider the neural network in the simplest case is actually layer after layer of spectrums of increasingly complex and more nebulous meaning stacked upon each other. With each spectrum ranging from fully deterministic through chaotic to random[1].
In effect the trigger is a match to a characteristic on one or more of the spectrums.
If the trigger is sufficiently different then it becomes easier to spot under any kind of analysis we might be capable of doing in some future time.
It’s important to note “future time” as currently even with simple neural networks “reversing the weight” to something meaningful is not to dissimilar to trying to “reverse a one way function”. That is the NN either has to be weak in some way, or you have to go through every type of input state.
This becomes even more complex when you consider the spectrums are not built by humans but by the statistics of the input data and how it was entered (which is effectively unknown after the event).
One of the aims of “Hybrid AI” is to combine rule based “Expert Systems” with “Neural Networks” and “human guidance” to in effect add “mores, morals and ethics” to spot and correct issues in the neural network. How well this might work is yet to be determined, however there is a lot of political and social imperative, especially before AGI gets sufficient agency to do more harm than currently happens with “self driving vehicles” (which is currently way to much).
[1] I went through describing this just a few days back,
https://www.schneier.com/blog/archives/2024/01/chatbots-and-human-conversation.html/#comment-431491