We acquired 200 traces from the same setup. We noticed a desynchronization.
We used this synchronization function :
def min_dist_curve(curve,ref,shifting_window,pattern_window):
distmin = sys.maxint
for s in range(shifting_window[0],shifting_window[1]):
dist = 0
for p in range(pattern_window[0],pattern_window[1]):
dist = dist + np.abs(ref[p] - curve[p+s])
distmin = min(dist,distmin)
if distmin == dist:
offset = s
return offset,distmin After alignement we decided to change selection funtion from AES Sbox to AES AddRoundKey as we thought that masking has been implemented. We noticed that no jitter has been put around AddRoundKey function ... PERFECT.
Here is an overview after synchronization
Before 6500 we can see some fake AddRoundKey. The last set of AddRoundKey are locate betwee 6500 and 8500. After 8500 we can see the Sbox routines.
The script performing the attack looks like:
# Traces offset where ARK has been identified
tbegin = 6500
tend = 8500
byte = 0
found = 0
# Buffer containing the retrieved key
key = []
# Key space length
knumber = 256
# Hypothesis drived by key guess
hyp = np.zeros(shape=(256,len(textin)))
# We fill hypothesis for all messages given a key guess
for i in range(len(textin)):
for kg in range(knumber):
hyp[kg][i] = pysca.hw(kg^textin[i][byte])
# We go through with traces point per point
t = tbegin
while t < tend:
# Observation array containing all traces at time t
obs = []
for i in range(len(textin)):
obs.append(synctracedata[i][t])
# We test all possible key and compute a correlation
for kg in range(knumber):
# Correlation coefficient is included in [-1;1]
# Our setup have a reversed the power comsumption maybe to due a switch
# Nevermind the correlation is negative, an absolute value can be taken
# but it will bring couple of fake correlation
if (np.corrcoef(hyp[kg],obs)[0][1] < -0.7):
if not found:
print byte, t, hex(kg)
key.append(kg)
found = 1
break
if found and (np.abs(np.corrcoef(hyp[kg],obs)[0][1]) < 0.7):
found = 0
byte = byte + 1
t = t + 1 The attack retrieved all key bytes with some ambiguity. We had a plaintext and a ciphertext. So we bruteforced key bytes which were ambiguous. The good key has been then recovered.
Chipwhisperer CWAnalyzer data compressed: StillNotScary_CWAnalyzer_200traces_data.7z
Note: Those traces have been captured using HydraRHME2 Hardware + ChipWisperer-Lite Hardware