- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
AdaDoS is a proposed automated attack on SDNs to cause degradation of network throughput while minimizing the likelihood of detection by an IDS. It is trained using a reciprocal policy transfer, where a teacher agent receives the true state as input, and uses it to guide the reward shaping of a student model. AdaDoS was compared to LDoS and Double-LDoS, and they found that the attack success rate was much higher against both models, with similar average bandwidth compared to Double-LDoS (meaning, both models caused similar damage to the network, but AdaDoS was less likely to be caught). They also tested the model’s robustness to noise, but found that it causes substantial decrease in attack success rate, concluding this would be a successful defense strategy against their method.
Critical strengths
- The teacher-student reward shaping is a good idea for POMDPs, especially for automated security agents where observations can differ so broadly from the true states
- The results on the baselines are strong both in terms of ASR and bandwidth
- The theory is good and the proofs appear sound.
Critical weaknesses
- Is this really “adversarial RL”? The teacher and student models are trying to align their KL divergence. They are working in conjunction, not adversarially.
- There is no ablation of how well the student model learns without its teacher, which is a critical part of the claimed contributions of this work
- The threat model seems unrealistic to me. Would the attacker really have observability into the SDN and IDS?
- The reward function requires more justification. Why is the reward for congestion scaled by z-c^2?
- The noise ranges in Section 5.4 are very wide assuming the data is normalized. Especially with the mean at 10 and std at 5, the noise is probably dominating the signal. This experiment should be re-run with less noise to determine the boundary for when the additional noise causes issues, rather than showing 3 instances where it pushes the model back to essentially random behavior
- The brief discussion of GANs at the end seems out of place in the paper
- The conclusion needs to be rewritten. There are numerous typos, and references undefined terms (“survival ability problem”) that appear nowhere else
Nitpicks:
- SDN should be defined in the abstract and intro before it is used as an acronym
- Introduction says contributions are threefold, then lists 4 contributions
- The specific appendix with the detailed list of topologies is not mentioned
- Table 1 needs more significant figures in some of the results; the ‘/’ in the title of Table 2 never appears
- The two tables should also appear closer in the text to where they are referenced if possible
Questions for authors
- How does just the student model do without the help of the teacher’s transfer learning?
- How would the strong affect of noise shown in Section 5.4 affect deploying this model in the (noisy) real world?
- What was the motivation for the reward function, in particular the reward for congestion?