Тест по отравлению данных и бэкдор-атакам

Ответы и пояснения появляются по мере прохождения теста.

1. Which goal distinguishes a triggered backdoor from general availability poisoning?
   • Attacker-chosen behavior associated with a specified trigger
   • A requirement that every clean input must be misclassified
   • A guarantee that all triggered inputs succeed without errors
   • A requirement that only image models can be attacked
2. What does clean-label poisoning preserve in its modified training examples?
   • A guarantee that every example is harmless
   • Every original input byte with no change
   • A label appropriate to the training item
   • A guarantee that no defense can detect the poison
3. Что продемонстрировало исследование BadNets?
   • Более быстрый алгоритм обучения
   • Знак остановки ошибочно классифицирован как знак ограничения скорости из-за триггера-наклейки.
   • Новый оптимизатор
   • Способ сжатия моделей
4. Как исследователи показали, что наборы данных веб-масштаба можно дешево отравить?
   • Взлом облачных серверов
   • Подкуп аннотаторов
   • Перегружая сеть
   • Покупая домены с истекшим сроком действия, на которых размещена часть URL-адресов изображений набора данных.
5. What does Spectral Signatures in Backdoor Attacks study as a defense signal?
   • A proof that all larger batches eliminate triggers
   • Only document filenames sorted alphabetically
   • Suspicious statistical structure in learned representations
   • A guarantee that all clean accuracy equals zero
6. Why can a dataset index pointing to remote content have a split-view poisoning exposure?
   • The indexed resource can change before a later downloader retrieves it
   • Every URL necessarily points to an immutable archived copy
   • All dataset labels are automatically rewritten at inference
   • A domain’s first owner must keep permanent control forever
7. Which stage is directly manipulated in a data-poisoning attack, as opposed to an inference-only input perturbation?
   • Only the final user-interface font
   • Only a previously fixed output screenshot
   • Only the displayed model name
   • The training data or learning process
8. What reference makes a content hash useful against the cited split-view poisoning threat?
   • A fresh hash supplied solely by the attacker alongside the changed resource
   • A trusted hash of the raw resource obtained before the malicious change
   • Only an unchanged URL string
   • A requirement that the resource keep the same file extension