Day 17 · Tue 29 Sep · Steganography and Backdoors: Daily Quiz

Request edit access

JavaScript isn't enabled in your browser, so this file can't be opened. Enable and reload.

Day 17 · Tue 29 Sep · Steganography and Backdoors (E.2) · Stephan Wäldchen

This is a simple quiz to help you recall what you have learned today, and to confirm you have achieved the learning outcomes. The questions are deliberately easy. Passing this quiz is *not* a replacement for being in the course. You do not need to submit the quiz, and we will not grade it.

Nothing is collected about who you are. If you do submit, your score and the correct answers appear immediately.

What is steganography?

Encrypting a secret message

Hiding the exchange of a secret message

Fooling someone into reading the wrong message

Using pseudo-random functions

Clear selection

Which notion of security is stronger?

Information-theoretically secure steganography

Computationally secure steganography

Clear selection

Regarding the previous question, why do we nevertheless rely on the weaker notion of security?

Because it requires shorter cover messages

Because it requires shorter keys

Because it’s still secure against poly-time adversaries

Because it allows for easier decoding

Clear selection

What properties does a keyed pseudo-random function have? (Select all that apply.)

Passes all common randomness tests

Cannot be distinguished from a truly random function by a poly-time algorithm

It is truly deterministic

It is truly random

The key doesn’t influence the output

The function itself can be calculated efficiently

What makes a computationally-secure steganographic scheme successful? (Select all that apply.)

The decoder can recover the secret message

The cover message contains no information about the secret message

The adversary cannot recover the secret message

The adversary cannot distinguish the cover message from innocent messages

Even with infinite compute for the adversary the method remains secure

What is Merlin-Arthur classification?

An interactive, adversarial classification setup

A classification setup with student and teacher model

A classification method based on AI safety via debate

Clear selection

What are conditions for Merlin-Arthur interpretability to guarantee highly precise features? (Select all that apply.)

Morgana is at least as computationally strong as Merlin

Merlin is at least as computationally strong as Morgana

Arthur and Merlin classify with 100% accuracy

Arthur has high soundness (robustness against Morgana)

Arthur must be implemented as a neural network

It only applies to image models

What does “unelicitable” mean?

The backdoor is not white-box detectable

The backdoor isn’t black-box detectable

The trigger of the backdoor cannot be recovered

The behaviour on the trigger cannot be predicted

Clear selection

Why is it infeasible to guarantee that an interpretation is faithful? (Select all that apply.)

The brute force proof is very expensive itself

There are exponentially many paths that we need to track

Noise across layers in a path propagates badly

The interpretation has to be re-derived from scratch for every new input

Any interpretation is a lossy approximation, so it can never be exactly faithful by definition

The input space becomes very large

Steganography and backdoors are both about behaviour that is present but hard to see. Which of the two worries you more in a deployed system, and what would you want to be able to check before shipping it?

Submit

Clear form

Never submit passwords through Google Forms.

This form was created inside of Iliad.

Does this form look suspicious? Report

Forms