Lightweight Security-by-Design for TinyML Models in Constrained Embedded Environments

. 2026 Sep 9;26(18):5727. doi: 10.3390/s26185727

Algorithm 2 Adaptive Runtime Defense Monitor

1:
Input: query x, baseline distribution q, threshold $τ$ , window size n
2:
State: query window $W \leftarrow \emptyset$ , defense level $L \leftarrow 0$
3:
Add x to W
4:
if $| W | \geq n$ then
5:
Compute $p (t)$ from W
6:
Compute $D_{t}$ using Equation (5)
7:
if $D_{t} > τ$ then
8:
$L \leftarrow 1$
9:
Apply output rounding and rate limit
10:
else
11:
$L \leftarrow 0$
12:
end if
13:
Clear W
14:
end if
15:
return model output at defense level L