Mismatched Binary Hypothesis Testing: Error Exponent Sensitivity

We study the problem of mismatched binary hypothesis testing between i.i.d. distributions. analyze tradeoff pairwise error probability exponents when actual distributions generating observation are different from used in likelihood ratio test, sequential and Hoeffding’s generalized test composite setting. When real within a small divergence ball distributions, we find deviation worst-case exponent each with respect to matched exponent. In addition, consider case where an adversary tampers observation, again type. show that tests more sensitive distribution mismatch than adversarial tampering.