Iran Protests: DPI blocking of Instagram (Part 2)

In early January 2018, OONI published a post reporting on the blocking of Telegram and Instagram amidst protests in Iran. We have since been analyzing RIPE data and other network measurements collected from Iran in an attempt to better understand the blockages. Upon further analysis, we found that Instagram was in fact blocked (during the Iran protests) through the use of Deep Packet Inspection (DPI) technology which targeted the TLS protocol. And it was configured in an interesting way.

TLS is fun

First, most Iranian RIPE Atlas probes failed to do TLS handshake with Instagram, but TCP connection was established with reasonable TTC (time-to-connect).

Second, IP addresses that were collected using Iranian DNS resolvers were giving proper TLS certificate for, which meant that DNS was not tampered with.

Third, Iranian probes were able to TCP-connect to endpoints collected with “external” resolvers, but failed to do TLS handshake (besides single non-routable IP address that came as a cached DNS response from some resolver in Harbin, China).

Fourth, TLS cert measurement to “nearby” IP address was successful, which meant that the whole AS was not blackholed.

Fifth, TLS cert measurement with bogus SNI field (or without SNI field at all) was blocked too. And so was a request to a “nearby” IP address with Instagram SNI, which meant that the SNI field is inspected, but it’s not the only one. Moreover, a request to a completely unrelated host with SNI header was blocked.

And now the fun starts: If DPI content inspection is applied to all traffic, then it’s possible to create a self-signed certificate for and have some fun as DPI does not validate certificate. It’s possible to send a certificate for an absolutely unrelated SNI request, see the response blocked and deduce that commonName field of the certificate that is sent in clear-text in response is used for filtering. It’s possible to check TLS over non-standard port and deduce that it’s not only TCP port #443 that is inspected with DPI.

But more fun can be achieved with RIPE Atlas and some packet crafting with scapy and nfqueue. For example, one can intercept a packet with ServerHello, override commonName on the wire (look at awesome PKI Posters if you don’t want to read RFC on packet structure) and make parasitic traceroute out of that with commonName displayed in RIPE Atlas measurement being set to TTL of the last hop! Sending packets is easy, but you’ll have to look at pcap if you want to know actual traceroute information:

def ip_ttl_eq_ssl_cn(payload, data):
    bag = []
    for i in xrange(32):
        pkt = IP(data)
        pkt[IP].ttl = i
        del pkt[IP].chksum
        pkt[TCP].payload = str(pkt[TCP].payload).replace(
                '', 'in{:03d}gram.rus'.format(i))
        del pkt[TCP].chksum

More practical results may be achieved splitting a single ServerHello segment into smaller ones at “bad word” boundary. That’s basically the same behaviour that brdgrd and GoodbyeDPI trigger by manipulating TCP Window. And that shows that lots of probes bypass commonName filter when commonName is split into several TCP segments using the following code:

def tcp_segmentation(payload, data):
    pkt = IP(data)
    segment = str(pkt[TCP].payload)
    slices = []
    badword = ''
    while badword in segment:
        ndx = segment.index(badword)
        segment = segment[ndx+5:]
    bag = []
    offset = 0
    for p in slices:
        pkt = IP(data)
        pkt[TCP].payload = p
        pkt[TCP].seq += offset
        del pkt[IP].len, pkt[IP].chksum, pkt[TCP].chksum
        offset += len(p)

So my conclusion is that it’s DPI filtering traffic (not a transparent proxy) as packets are not reassembled back into the TCP stream. That also shows that both SNI and commonName are inspected, so both traffic flows (ingress and egress) are passed to inspection with all the performance implications of alike deployment.

Unfortunately, it’s hard to tell the reason for connect() failures, but it seems that a network congestion (or shutdown) might have taken place during the event. The amount of RIPE Atlas probes going offline in that time-frame makes me think so. The quantitative analysis of the collected data is yet to be done.

You can get sample python code doing server-side packet manipulation in

Improving OONI DNS heuristics

OONI data confirmed the blocking of Instagram during the latest protests in Iran, but the means of censorship was misidentified. It looked like Instagram was blocked by means of DNS tampering, but upon further analysis, that appears to have been a false positive because:

Limitations to OONI’s Web Connectivity test (and how they can result in false positives) are documented on the OONI website. Unfortunately, there are many cases where DNS-related false positives can emerge, and we quite recently reported on cases of DNS misconfiguration that looked like censorship, but weren’t.

This highlights the need to improve our heuristics on accurately identifying cases of DNS-based censorship. We have identified some heuristics for detecting cases of DNS hijacking, and we hope to implement them as OONI Probe tests within the next year. In the meanwhile, we encourage you to reach out to us to share knowledge and ideas you may have for better detecting cases of DNS-based censorship!